Daily Papers

Traditional language model-based theorem proving assumes that by training on a sufficient amount of formal proof data, a model will learn to prove theorems. Our key observation is that a wealth of informal information that is not present in formal proofs can be useful for learning to prove theorems. For instance, humans think through steps of a proof, but this thought process is not visible in the resulting code. We present Lean-STaR, a framework for training language models to produce informal thoughts prior to each step of a proof, thereby boosting the model's theorem-proving capabilities. Lean-STaR uses retrospective ground-truth tactics to generate synthetic thoughts for training the language model. At inference time, the trained model directly generates the thoughts prior to the prediction of the tactics in each proof step. Building on the self-taught reasoner framework, we then apply expert iteration to further fine-tune the model on the correct proofs it samples and verifies using the Lean solver. Lean-STaR achieves state-of-the-art results on the miniF2F-test benchmark within the Lean theorem proving environment, significantly outperforming base models (43.4% rightarrow 46.3%, Pass@64). We also analyze the impact of the augmented thoughts on various aspects of the theorem proving process, providing insights into their effectiveness.

Coding agents now produce a growing share of a team's code, while the reasoning behind each change -- the alternatives weighed, the constraints discovered, the approaches rejected -- is trapped in assistant transcripts that vanish with the session. Memory for this setting, the agentic development lifecycle (ADLC), is usually posed as one retrieval problem and built as machinery: tiered stores, memory graphs, compiled wikis, model-judged admission. We argue memory should instead be git-bound -- built into the repository's version control, inheriting the guarantees the machinery struggles to construct: ground truth from commits, freshness from rebuild, verification from the merge, containment from review. On this ledger we solve two problems separately, then combine them. Seed supply is closed as an eight-corpus retrieval study under a pre-registered ship discipline: five imported ranking mechanisms rejected, two kept, and a best configuration of ~0.31 pooled MRR -- ~60x the raw-transcript grep floor, ~15x an honest parsed-turn floor. Answer assembly is where ranking stops helping: single-shot retrieval scores only 0.07-0.20 answer-sufficiency on real developer questions, and ungated episode injection measurably degrades good answers. A router dispatches breadth to a git-anchored structural map, pointed lookups to confidence-gated episodes, and rationale to decision synthesis, which reconstructs why-arcs no single session contains (0.83 sufficiency on a young ~50k-LOC production system). Routed, the system answers at 382-980 tokens per question -- three orders of magnitude below the recorded history. Because ground truth is mined from commit-session links rather than annotated, every result is replicable on any user's own history at zero labeling cost. The remaining constraint is capture. Code, benchmark, and paper source: github.com/rekal-dev/rekal-cli.

LLM agents act in external environments where each action changes the state that later decisions condition on, and where a single wrong step can waste interaction budget or trigger irreversible side effects long before the final failure is observed. Reliable deployment therefore requires step-level confidence estimation: a calibrated probability that each proposed action is productive, available before the action is executed. Existing LLM confidence estimators are designed to score a response from the given prompt, but agent confidence also depends on execution consequences: whether similar actions in similar situations actually advanced the task after the environment responded. We introduce the \method (\methodshort), a self-evolving critic framework in which an LLM critic accumulates evidence from its own past judgments and their observed consequences. After each trajectory, a hindsight LLM that sees the full execution feedback votes on whether each step was productive. The resulting pseudo-labels populate a memory bank from which related productive and unproductive experiences are retrieved into the critic's prompt whenever a similar step recurs. \methodshort requires no training and uses no ground truth step labels. Across three agent benchmarks and three critic backbones, \methodshort attains the best calibration (ECE and Brier) and ranking (AUC) in every dataset--critic combination, reducing ECE by up to 54% relative to the strongest training-free baseline.

The acceleration in the adoption of AI-based automated decision-making systems poses a challenge for evaluating the fairness of algorithmic decisions, especially in the absence of ground truth. When designing interventions, uplift modeling is used extensively to identify candidates that are likely to benefit from treatment. However, these models remain particularly susceptible to fairness evaluation due to the lack of ground truth on the outcome measure since a candidate cannot be in both treatment and control simultaneously. In this article, we propose a framework that overcomes the missing ground truth problem by generating surrogates to serve as a proxy for counterfactual labels of uplift modeling campaigns. We then leverage the surrogate ground truth to conduct a more comprehensive binary fairness evaluation. We show how to apply the approach in a comprehensive study from a real-world marketing campaign for promotional offers and demonstrate its enhancement for fairness evaluation.

Various test-time interventions for Computer Use Agents (CUAs), including critic models, have been developed to improve performance through pre-execution action evaluation in complex Graphical User Interface (GUI) environments. However, existing critics suffer from two key limitations: they (1) focus primarily on short-sighted decision loops (e.g., forgetting earlier actions) and (2) lack the visual grounding needed to detect flawed actions (e.g., clicking wrong UI elements). To address these, we introduce HiViG, a History-aware Visually Grounded test-time framework, built around a multimodal critic trained on real GUI trajectories to abstract past interactions into a compact record and to evaluate actions with visual grounding. At test time, HiViG integrates the critic into the policy decision loop to provide macro-action history, which summarizes the policy's completed achievements, and visually grounded critique, which verifies raw execution coordinates against the current screenshot to intercept errors before execution. Across web, mobile, and desktop benchmarks, HiViG consistently outperforms existing scalar and verbal critics, improving average success rates over the strongest baseline by 5.8% for Qwen3-VL-32B and 9.0% for Gemini-3-Flash, and demonstrates strong cross-platform generalization. Ablations show that macro-action history mitigates short-sighted planning and visually grounded critique reduces execution errors, with both components being critical for test-time scaling in long-horizon GUI tasks.

Tool-Integrated Reasoning (TIR) with search engines enables large language models to iteratively retrieve up-to-date external knowledge, enhancing adaptability and generalization in complex question-answering tasks. However, existing search agent pipelines typically depend on reinforcement learning based optimization, which often suffers from sparse outcome rewards, leading to inefficient exploration and unstable training. We introduce CriticSearch, a fine-grained credit-assignment framework that supplies dense, turn-level feedback via a retrospective critic mechanism. During training, a frozen, asymmetric critique LLM retrospectively evaluates each turn using privileged information from the full trajectory and gold answers, converting these assessments into stable, dense rewards that guide policy improvement. Experimental results across diverse multi-hop reasoning benchmarks demonstrate that CriticSearch consistently outperforms existing baselines, achieving faster convergence, improved training stability, and higher performance.

AI agents deployed as persistent assistants must maintain correct beliefs as their information environment evolves. In practice, evidence is scattered across heterogeneous sources that often contradict one another, new information can invalidate earlier conclusions, and user preferences surface through corrections rather than explicit instructions. Existing benchmarks largely assume static, single-authority settings and do not evaluate whether agents can keep up with this complexity. We introduce ClawArena, a benchmark for evaluating AI agents in evolving information environments. Each scenario maintains a complete hidden ground truth while exposing the agent only to noisy, partial, and sometimes contradictory traces across multi-channel sessions, workspace files, and staged updates. Evaluation is organized around three coupled challenges: multi-source conflict reasoning, dynamic belief revision, and implicit personalization, whose interactions yield a 14-category question taxonomy. Two question formats, multi-choice (set-selection) and shell-based executable checks, test both reasoning and workspace grounding. The current release contains 64 scenarios across 8 professional domains, totaling 1{,}879 evaluation rounds and 365 dynamic updates. Experiments on five agent frameworks and five language models show that both model capability (15.4% range) and framework design (9.2%) substantially affect performance, that self-evolving skill frameworks can partially close model-capability gaps, and that belief revision difficulty is determined by update design strategy rather than the mere presence of updates. Code is available at https://github.com/aiming-lab/ClawArena.

Can AI reason about a war before its trajectory becomes historically obvious? Analyzing this capability is difficult because retrospective geopolitical prediction is heavily confounded by training-data leakage. We address this challenge through a temporally grounded case study of the early stages of the 2026 Middle East conflict, which unfolded after the training cutoff of current frontier models. We construct 11 critical temporal nodes, 42 node-specific verifiable questions, and 5 general exploratory questions, requiring models to reason only from information that would have been publicly available at each moment. This design substantially mitigates training-data leakage concerns, creating a setting well-suited for studying how models analyze an unfolding crisis under the fog of war, and provides, to our knowledge, the first temporally grounded analysis of LLM reasoning in an ongoing geopolitical conflict. Our analysis reveals three main findings. First, current state-of-the-art large language models often display a striking degree of strategic realism, reasoning beyond surface rhetoric toward deeper structural incentives. Second, this capability is uneven across domains: models are more reliable in economically and logistically structured settings than in politically ambiguous multi-actor environments. Finally, model narratives evolve over time, shifting from early expectations of rapid containment toward more systemic accounts of regional entrenchment and attritional de-escalation. Since the conflict remains ongoing at the time of writing, this work can serve as an archival snapshot of model reasoning during an unfolding geopolitical crisis, enabling future studies without the hindsight bias of retrospective analysis.

Reasoning traces from chain-of-thought models appear to offer a legible window into how a model arrives at its answer. A growing body of work treats them as such, using LLM judges to diagnose errors, evaluate faithfulness, and provide step-level supervision via process reward models and generative critics. These practices rely on the text of a reasoning step carrying information about its functional role. But does the text actually encode information about which reasoning steps matter? We operationalize the importance of a reasoning step as its advantage: the change in expected reward, e.g., producing the correct final answer, from including that step, estimated via Monte Carlo rollouts. Basing ground truth on these estimates, we evaluate whether LLM judges can identify high-advantage steps and find that sufficiently capable LLMs can outperform a prevalence baseline but fall well short of a noise ceiling. Fine-tuning a model as a step-level critic yields strong improvement for incorrect responses but remains distant from ceiling for correct responses, suggesting that step importance is only partially recoverable from the text of the reasoning trace. Our findings contribute to a growing body of chain-of-thought faithfulness work that cautions against treating the legibility of reasoning traces as interpretability, especially with implications for process reward modeling.

Reinforcement learning for long-horizon agents relies on purely retrospective training signals: credit is assigned only after observing environmental consequences, leaving the agent's belief at action time invisible to the gradient. We introduce Prospective Hindsight (PH), a self-calibrating training principle that augments any retrospective base method with a signal derived from the gap between the agent's prospective prediction (before feedback) and the retrospective evaluation (after feedback). This per-rollout surprise identifies samples where the agent's self-model is most inaccurate and amplifies their gradient contribution through a stop-gradient surprise-weighted advantage. Since the prospective predictor shares parameters with the policy, the two co-evolve, progressively shifting focus to the agent's remaining blind spots. We connect this principle to a privileged-information gap and show that minimizing the surprise residual provides a descent pathway on the agent's miscalibration rate; calibration thus emerges as a byproduct of optimization rather than from an added objective. On single-turn verifiable tasks and a multi-turn personal-agent task (under GRPO, on-policy distillation, and their combination), PH improves both task performance and calibration, with consistent gains across model scales. Notably, the dominant miscalibration mode shifts structurally between regimes, overconfident failures in single-turn, underconfident successes in multi-turn, yet the same training principle addresses both successfully.

Large Language Model (LLM) agents are increasingly expected to maintain coherent, long-term personalized memory, yet current benchmarks primarily measure static fact retrieval, overlooking the ability to revise stored beliefs when new evidence emerges. We identify a critical and underexplored failure mode, Implicit Conflict: a later observation invalidates an earlier memory without explicit negation, requiring contextual inference and commonsense reasoning to detect. To rigorously evaluate this capability, we introduce STALE, a benchmark of 400 expert-validated conflict scenarios (1,200 evaluation queries across three probing dimensions) spanning over 100 everyday topics with contexts up to 150K tokens. We propose a three-dimensional probing framework that tests State Resolution (detecting that a prior belief is outdated), Premise Resistance (rejecting queries that falsely presuppose a stale state), and Implicit Policy Adaptation (proactively applying updated states in downstream behavior). A systematic evaluation of frontier LLMs and specialized memory frameworks reveals a pervasive gap between retrieving updated evidence and acting on it, with even the best evaluated model achieving only 55.2% overall accuracy. Models often accept outdated assumptions embedded in a user's query, and they struggle to recognize when a change in one aspect of the user's state should invalidate related memories. To establish an initial baseline for state-aware memory, we further present CUPMem, a prototype that strengthens write-time revision through structured state consolidation and propagation-aware search, suggesting that explicit state adjudication is a promising direction for robust agentic memory.

Automated fact-checking remains a challenge for Large Language Models (LLMs) due to "query brittleness" in traditional retrieval systems. We propose DeLIVeR (Decomposed Learning for Information-grounded Veracity Recognition), a framework that treats evidence retrieval as a reinforced strategic exploration task. DeLIVeR utilizes a Planner LLM to decompose complex claims into targeted question sets, which are used to traverse structured Knowledge Graphs (KGs) for high-precision evidence. We optimize the Planner's policy using Group Relative Policy Optimization (GRPO) with a reward system prioritizing structural diversity and verdict accuracy. Our evaluation on LIAR, FEVER, and PolitiFact shows that DeLIVeR significantly outperforms state-of-the-art baselines. Using Qwen2.5-7B, our framework achieved peak F1-scores of 83.73, 84.57, and 79.70 respectively, representing a 10-15% improvement over HippoRAG2. By shifting to a reinforced question-planning strategy, DeLIVeR effectively bridges multi-hop reasoning gaps and provides an auditable, transparent path for verifiable misinformation detection.

Fake news and misinformation poses a significant threat to society, making efficient mitigation essential. However, manual fact-checking is costly and lacks scalability. Large Language Models (LLMs) offer promise in automating counter-response generation to mitigate misinformation, but a critical challenge lies in their tendency to hallucinate non-factual information. Existing models mainly rely on LLM self-feedback to reduce hallucination, but this approach is computationally expensive. In this paper, we propose MisMitiFact, Misinformation Mitigation grounded in Facts, an efficient framework for generating fact-grounded counter-responses at scale. MisMitiFact generates simple critique feedback to refine LLM outputs, ensuring responses are grounded in evidence. We develop lightweight, fine-grained critique models trained on data sourced from readily available fact-checking sites to identify and correct errors in key elements such as numerals, entities, and topics in LLM generations. Experiments show that MisMitiFact generates counter-responses of comparable quality to LLMs' self-feedback while using significantly smaller critique models. Importantly, it achieves ~5x increase in feedback generation throughput, making it highly suitable for cost-effective, large-scale misinformation mitigation. Code and LLM prompt templates are at https://github.com/xxfwin/MisMitiFact.

Eliciting detailed and actionable software requirements from end-users is a critical phase in the iterative development of a software product or application. To ensure the feedback collected is detailed and actionable, software teams can leverage the laddering interview technique. While effective for ensuring granular and actionable items from the software feedback, these interviews are subject to several limitations. They are traditionally a manual process associated with a time and financial burden, limiting scalability; interviewers must balance probing for depth while managing interviewee behavioral and cultural constraints. To address these limitations, we present LadderTeam, an open, reproducible framework that automates UX wireframe interviews using a dual-agent Large Language Model (LLM) architecture. An active interviewer agent executes one of three probing strategies (ACV, 5-Whys, and JTBD) to elicit actionable software requirements from usability feedback comments, while a concurrent background Judge agent evaluates probe-response pairs and triggers real-time guardrails to prevent topic drift. To rigorously evaluate LLM laddering without participant variance confounds, we introduce a controlled simulation methodology utilizing scripted ground-truth transcripts to isolate probe quality as the sole experimental variable. Across 216 interviews, LadderTeam achieved 99.1\% chain convergence and an 81.0\% ground-truth actionable response match (86.1\% reluctant personality, 75.9\% terse personality) with zero drift across all runs. All evaluation code, all transcripts, inputs, and a live demonstration platform will be open-sourced upon acceptance.

AI agents are being increasingly deployed in dynamic, open-ended environments that require adapting to new information as it arrives. To efficiently measure this capability for realistic use-cases, we propose building grounded simulations that replay real-world events in the order they occurred. We build FutureSim, where agents forecast world events beyond their knowledge cutoff while interacting with a chronological replay of the world: real news articles arriving and questions resolving over the simulated period. We evaluate frontier agents in their native harness, testing their ability to predict world events over a three-month period from January to March 2026. FutureSim reveals a clear separation in their capabilities, with the best agent's accuracy being 25%, and many having worse Brier skill score than making no prediction at all. Through careful ablations, we show how FutureSim offers a realistic setting to study emerging research directions like long-horizon test-time adaptation, search, memory, and reasoning about uncertainty. Overall, we hope our benchmark design paves the way to measure AI progress on open-ended adaptation spanning long time-horizons in the real world.

Interactive language-model agents increasingly solve complex tasks through long-horizon, multi-call reasoning, where errors in beliefs or actions can compound across tool interactions. Retrospective search can recover from such failures but is prone to misallocation. Delayed outcomes obscure the contribution of intermediate search decisions, leading to Attribution Complexity, while evolving execution evidence leads to Adaptation Complexity, where previously learned estimates become stale. To address these challenges, we first introduce Search Value Dynamics (SVD), which characterizes the evolving trade-off between the gain and cost of retrospective search. Building on SVD, we propose ForkPilot, a self-evolving two-stage policy-learning framework. In the first stage, ForkPilot learns a search-value policy offline from completed trajectories through automatically constructed outcome comparisons. In the second stage, it makes search decisions based on current observations and then self-evolves by incorporating newly completed trajectories into subsequent policy updates. We evaluate ForkPilot across 6 diverse benchmarks and 7 widely used LLM backbone families, including four open-source families, GPT-5.6 Sol, and Opus 4.8 in a production agentic system, against 9 competitive baselines, including a real-world harness deployment used by hundreds of thousands of paid users. ForkPilot achieves comparable state-of-the-art performance while reducing token usage by up to 59.2%, demonstrating its efficacy.

The rapid propagation of misinformation poses substantial risks to public interest. To combat misinformation, large language models (LLMs) are adapted to automatically verify claim credibility. Nevertheless, existing methods heavily rely on the embedded knowledge within LLMs and / or black-box APIs for evidence collection, leading to subpar performance with smaller LLMs or upon unreliable context. In this paper, we propose retrieval augmented fact verification through the synthesis of contrasting arguments (RAFTS). Upon input claims, RAFTS starts with evidence retrieval, where we design a retrieval pipeline to collect and re-rank relevant documents from verifiable sources. Then, RAFTS forms contrastive arguments (i.e., supporting or refuting) conditioned on the retrieved evidence. In addition, RAFTS leverages an embedding model to identify informative demonstrations, followed by in-context prompting to generate the prediction and explanation. Our method effectively retrieves relevant documents as evidence and evaluates arguments from varying perspectives, incorporating nuanced information for fine-grained decision-making. Combined with informative in-context examples as prior, RAFTS achieves significant improvements to supervised and LLM baselines without complex prompts. We demonstrate the effectiveness of our method through extensive experiments, where RAFTS can outperform GPT-based methods with a significantly smaller 7B LLM.

With the emergence of large-scale vision language models (VLM), it is now possible to produce realistic-looking radiology reports for chest X-ray images. However, their clinical translation has been hampered by the factual errors and hallucinations in the produced descriptions during inference. In this paper, we present a novel phrase-grounded fact-checking model (FC model) that detects errors in findings and their indicated locations in automatically generated chest radiology reports. Specifically, we simulate the errors in reports through a large synthetic dataset derived by perturbing findings and their locations in ground truth reports to form real and fake findings-location pairs with images. A new multi-label cross-modal contrastive regression network is then trained on this dataset. We present results demonstrating the robustness of our method in terms of accuracy of finding veracity prediction and localization on multiple X-ray datasets. We also show its effectiveness for error detection in reports of SOTA report generators on multiple datasets achieving a concordance correlation coefficient of 0.997 with ground truth-based verification, thus pointing to its utility during clinical inference in radiology workflows.

In many domains of argumentation, people's arguments are driven by so-called attitude roots, i.e., underlying beliefs and world views, and their corresponding attitude themes. Given the strength of these latent drivers of arguments, recent work in psychology suggests that instead of directly countering surface-level reasoning (e.g., falsifying given premises), one should follow an argumentation style inspired by the Jiu-Jitsu 'soft' combat system (Hornsey and Fielding, 2017): first, identify an arguer's attitude roots and themes, and then choose a prototypical rebuttal that is aligned with those drivers instead of invalidating those. In this work, we are the first to explore Jiu-Jitsu argumentation for peer review by proposing the novel task of attitude and theme-guided rebuttal generation. To this end, we enrich an existing dataset for discourse structure in peer reviews with attitude roots, attitude themes, and canonical rebuttals. To facilitate this process, we recast established annotation concepts from the domain of peer reviews (e.g., aspects a review sentence is relating to) and train domain-specific models. We then propose strong rebuttal generation strategies, which we benchmark on our novel dataset for the task of end-to-end attitude and theme-guided rebuttal generation and two subtasks.

The proliferation of online misinformation has posed significant threats to public interest. While numerous online users actively participate in the combat against misinformation, many of such responses can be characterized by the lack of politeness and supporting facts. As a solution, text generation approaches are proposed to automatically produce counter-misinformation responses. Nevertheless, existing methods are often trained end-to-end without leveraging external knowledge, resulting in subpar text quality and excessively repetitive responses. In this paper, we propose retrieval augmented response generation for online misinformation (RARG), which collects supporting evidence from scientific sources and generates counter-misinformation responses based on the evidences. In particular, our RARG consists of two stages: (1) evidence collection, where we design a retrieval pipeline to retrieve and rerank evidence documents using a database comprising over 1M academic articles; (2) response generation, in which we align large language models (LLMs) to generate evidence-based responses via reinforcement learning from human feedback (RLHF). We propose a reward function to maximize the utilization of the retrieved evidence while maintaining the quality of the generated text, which yields polite and factual responses that clearly refutes misinformation. To demonstrate the effectiveness of our method, we study the case of COVID-19 and perform extensive experiments with both in- and cross-domain datasets, where RARG consistently outperforms baselines by generating high-quality counter-misinformation responses.

Counterfactual Risk Minimization (CRM) is a framework for dealing with the logged bandit feedback problem, where the goal is to improve a logging policy using offline data. In this paper, we explore the case where it is possible to deploy learned policies multiple times and acquire new data. We extend the CRM principle and its theory to this scenario, which we call "Sequential Counterfactual Risk Minimization (SCRM)." We introduce a novel counterfactual estimator and identify conditions that can improve the performance of CRM in terms of excess risk and regret rates, by using an analysis similar to restart strategies in accelerated optimization methods. We also provide an empirical evaluation of our method in both discrete and continuous action settings, and demonstrate the benefits of multiple deployments of CRM.

GUI agents are deployed with frozen weights and discard everything they experience on the job. Existing ways to update an agent's weights assume something deployment withholds: ground truth, rollouts beyond the single attempt (retries, samples, practice runs), or a learning phase other than deployment. Because GUI actions can be irreversible, a deployed agent gets one attempt per task occurrence, in arrival order, and every attempt counts. No ground truth is available at any point. We define fully test-time adaptation for GUI agents by these constraints and pair it with a minimal weight-space method, SOLO. Auxiliary models read each episode: a judge selects the episodes it deems successful, and a proposer-verifier pair relabels a failed episode's prefix with the subtask that prefix completed. Admitted episodes enter a short sliding window, and each admission updates a small adapter by top-K self-distillation on the agent's own predictions, provided the window holds a judged success. On recurring task streams built from WebArena, VisualWebArena and MobileWorld, SOLO improves on the frozen agent with both UI-TARS-7B and Qwen3-VL-8B, by three to six points of success rate, and exceeds two in-setting memory methods on the web streams.

Counterspeech effectively neutralizes the impact of online hate. Although prior work explores automated counterspeech generation, it largely emphasizes stylistic control while treating hate speech as homogeneous, overlooking that distinct forms of abuse require fundamentally different counterspeech strategies. To address this gap, we introduce FIRE (Factuality Informed Multi-Agent Reasoning Framework) that first decomposes hate speech into one of the five distinct categories (misinformation, stereotype, conspiracy, dehumanizing, non-factual), and then maps it to a targeted counterspeech style. To facilitate FIRE, we curate FactualCS, a novel dataset of 4,784 instances that provides the annotations regarding hate categories, reasoning traces, and evidence mappings, which are critical elements for grounded generation that are missing in prior work. A comprehensive evaluation across 28 baseline configurations demonstrates that FIRE significantly surpasses existing methods, despite using compact agents (<2B). FIRE achieves a sim 12 % and sim 11 % improvements in factual and category-specific accuracy respectively, while simultaneously reducing toxicity by sim 11 % relative to the strongest baselines. Further human evaluation confirms that responses generated by FIRE are significantly preferred over the strongest baselines, underscoring its effectiveness for real-world deployment. These findings show that decomposing the underlying intent of hate speech is essential for generating safe, effective, and contextually precise counterspeech.

Before letting an agent operate over real context, can you prove it used the right evidence? GroundEval turns that question into a deterministic test of what the agent searched, fetched, cited, and was permitted to access. In one case study, two frontier LLM judges scored a plausible agent response above 0.85. But the trace told a different story: the agent had never retrieved the artifact its answer depended on, yielding a GroundEval score of 0.000. We introduce GroundEval, a judge-free framework for evaluating agents against grounded, time-bounded, and access-controlled evidence. GroundEval uses a domain configuration to generate questions, lets the agent choose how to answer, and then scores both the final answer and the recorded trajectory that produced it. The benchmark targets three failures that LLM-as-judge evaluation struggles to detect: whether an agent checked before claiming absence, reasoned only from evidence available to the actor at the relevant time, and used the correct causal mechanism rather than a plausible one. These correspond to three tracks: Silence, Perspective, and Counterfactual. GroundEval exposes when plausible answers rest on invalid evidence paths, and produces structured per-question diagnostics that pair tool activity with the agent's turn-level narration, making each score inspectable rather than merely reported. What our case studies turned up is that this gap isn't some rare corner case. It's exactly the blind spot that final-answer and judge-based scoring were never built to catch.

Adversarial attacks in reinforcement learning (RL) often assume highly-privileged access to the victim's parameters, environment, or data. Instead, this paper proposes a novel adversarial setting called a Cheap Talk MDP in which an Adversary can merely append deterministic messages to the Victim's observation, resulting in a minimal range of influence. The Adversary cannot occlude ground truth, influence underlying environment dynamics or reward signals, introduce non-stationarity, add stochasticity, see the Victim's actions, or access their parameters. Additionally, we present a simple meta-learning algorithm called Adversarial Cheap Talk (ACT) to train Adversaries in this setting. We demonstrate that an Adversary trained with ACT still significantly influences the Victim's training and testing performance, despite the highly constrained setting. Affecting train-time performance reveals a new attack vector and provides insight into the success and failure modes of existing RL algorithms. More specifically, we show that an ACT Adversary is capable of harming performance by interfering with the learner's function approximation, or instead helping the Victim's performance by outputting useful features. Finally, we show that an ACT Adversary can manipulate messages during train-time to directly and arbitrarily control the Victim at test-time. Project video and code are available at https://sites.google.com/view/adversarial-cheap-talk

Generating accurate and calibrated confidence estimates is critical for deploying LLMs in high-stakes or user-facing applications, and remains an open challenge. Prior research has often framed confidence as a problem of eliciting a model's "self-knowledge", i.e., the ability of an LLM to judge whether its own answers are correct; this approach implicitly assumes that there is some privileged information about the answer's correctness that is accessible to the model itself. However, our experiments reveal that an LLM attempting to predict the correctness of its own outputs generally performs no better than an unrelated LLM. Moreover, we hypothesize that a key factor in building a "Correctness Model" (CM) is exposure to a target model's historical predictions. We propose multiple methods to inject this historical correctness information, creating a Generalized Correctness Model (GCM). We first show that GCMs can be trained on the correctness data from many LLMs and learn patterns for correctness prediction applicable across datasets and models. We then use CMs as a lens for studying the source of correctness prediction ability and its generalization, systematically controlling their training data and finding that answer phrasing is a strong predictor for correctness. We further explore alternative methods of injecting history without training an LLM, finding that including history as in-context examples can help improve correctness prediction, and post-hoc calibration can provide complementary reductions in calibration error. We evaluate GCMs based on Qwen3-8B across 5 model families and the MMLU and TriviaQA datasets, as well as on a downstream selective prediction task, finding that reliable LLM confidence estimation is a generalizable and model-agnostic skill learned by systematically encoding correctness history rather than a model-specific skill reliant on self-introspection.

Large Language Models have demonstrated remarkable capabilities across diverse tasks, yet they frequently generate hallucinations outputs that are fluent but factually incorrect or unsupported. We propose Counterfactual Probing, a novel approach for detecting and mitigating hallucinations in LLM outputs. Our method dynamically generates counterfactual statements that appear plausible but contain subtle factual errors, then evaluates the model's sensitivity to these perturbations. We hypothesize that genuine knowledge exhibits robustness to counterfactual variations, while hallucinated content shows inconsistent confidence patterns when confronted with plausible alternatives. Our comprehensive evaluation on TruthfulQA, factual statement datasets, and curated hallucination examples demonstrates that counterfactual probing achieves superior detection performance compared to baseline methods, while our adaptive mitigation strategies reduce hallucination scores by an average of 24.5%. The approach requires no model retraining and can be integrated into existing LLM pipelines as a realtime verification mechanism.

We present a framework for evaluating data-science agents on live, continuously updated data using executable ground truth and format-agnostic factoid scoring. Consider this example query: "what were last week's audience sizes"---the reference answer changes as the underlying data changes, so static references become outdated and standard LLM-as-a-judge pipelines cannot verify responses against a fixed ground truth. Our central contribution, ground-truth-as-code, encodes each expected answer as an executable reference function that recomputes the answer directly from live data at evaluation time, ensuring the reference remains consistent with the system it describes. We combine this with a factoid-level, format-agnostic judge that decomposes both the agent's response and the computed ground truth into atomic claims and scores precision, recall, and accuracy over them, irrespective of the response format (prose, list, table, HTML, etc.). The approach is applicable to agents whose expected outputs can be expressed as executable data computations. We validate the framework through a human--LLM agreement study on an internally developed machine learning skill deployed in production, using a synthetic database constructed to reproduce production schemas and entity relationships. Relative to a natural-language ground-truth baseline, our method achieves a 29% improvement in the Matthews Correlation Coefficient (MCC)---a class-balanced measure of agreement between expert annotators and LLM-as-a-judge predictions---and a 16% reduction in token consumption per test case, while a self-directed baseline lacking explicit ground truth is anti-correlated with human judgment. Agents that perform multi-source data integration and computation over non-stationary data are routinely deployed in industry; we propose ground-truth-as-code as a practical methodology for their evaluation.

Incrementally fine-tuning foundational models on new tasks or domains is now the de facto approach in NLP. A known pitfall of this approach is the catastrophic forgetting of prior knowledge that happens during fine-tuning. A common approach to alleviate such forgetting is to rehearse samples from prior tasks during fine-tuning. Several existing works assume a fixed memory buffer to store prior task examples, while relying on inferences (forward passes) with the model at hand for choosing examples for rehearsal from the buffer. However, given the increasing computational cost of model inference, and decreasing cost of data storage, we focus on the setting to rehearse samples with a fixed computational budget instead of a fixed memory budget. We propose a sampling scheme, \bf mix-cd, that prioritizes rehearsal of ``collateral damage'' samples, which are samples predicted correctly by the prior model but forgotten by the incrementally tuned one. The crux of our scheme is a procedure to efficiently estimate the density of collateral damage samples without incurring additional model inferences. Our approach is computationally efficient, easy to implement, and outperforms several leading continual learning methods in compute-constrained settings. All the code will be publicly available at https://github.com/jybai/mix-cd-rehearsal.

The spread of online misinformation threatens public health, democracy, and the broader society. While professional fact-checkers form the first line of defense by fact-checking popular false claims, they do not engage directly in conversations with misinformation spreaders. On the other hand, non-expert ordinary users act as eyes-on-the-ground who proactively counter misinformation -- recent research has shown that 96% counter-misinformation responses are made by ordinary users. However, research also found that 2/3 times, these responses are rude and lack evidence. This work seeks to create a counter-misinformation response generation model to empower users to effectively correct misinformation. This objective is challenging due to the absence of datasets containing ground-truth of ideal counter-misinformation responses, and the lack of models that can generate responses backed by communication theories. In this work, we create two novel datasets of misinformation and counter-misinformation response pairs from in-the-wild social media and crowdsourcing from college-educated students. We annotate the collected data to distinguish poor from ideal responses that are factual, polite, and refute misinformation. We propose MisinfoCorrect, a reinforcement learning-based framework that learns to generate counter-misinformation responses for an input misinformation post. The model rewards the generator to increase the politeness, factuality, and refutation attitude while retaining text fluency and relevancy. Quantitative and qualitative evaluation shows that our model outperforms several baselines by generating high-quality counter-responses. This work illustrates the promise of generative text models for social good -- here, to help create a safe and reliable information ecosystem. The code and data is accessible on https://github.com/claws-lab/MisinfoCorrect.

Standard Retrieval-Augmented Generation (RAG) systems predominantly rely on semantic relevance as a proxy for utility. However, this assumption collapses in realistic decision-making scenarios where user queries are laden with cognitive biases, such as false premises or confirmation bias. In such cases, maximizing relevance paradoxically promotes the retrieval of sycophantic evidence that reinforces hallucinations, a critical failure we term the ``Relevance-Robustness Gap''. To bridge this gap, we propose CoRM-RAG (Counterfactual Risk Minimization for RAG), a framework that aligns retrieval with decision safety rather than mere similarity. Grounded in causal intervention, we introduce a Cognitive Perturbation Protocol to simulate user biases during training, which is then distilled into a lightweight Evidence Critic. This scoring module learns to identify documents that possess sufficient evidential strength to steer the model toward correctness despite adversarial query perturbations. Extensive experiments on decision-making benchmarks demonstrate that CoRM-RAG significantly outperforms strong dense retrievers and LLM-based rerankers in adversarial settings, while enabling effective risk-aware abstention through reliable robustness scoring. Our code is available at https://github.com/PeiYangLiu/CoRM-RAG.git.

Vision-Language Navigation (VLN) requires agents to continuously ground task progress from long-horizon instructions and partial egocentric observations. Existing VLM-based navigation agents typically reason only over available observations and may remain confident even when task-relevant evidence is missing. For example, an agent may confidently proceed forward and get lost even though the landmark indicating the next turn lies outside its current field of view. We term this failure mode Progress Myopia: the agent fails to recognize unreliable progress grounding and continues acting on insufficient evidence. To address it, we propose SeekVLN, an evidence-seeking framework that couples semantic progress reasoning with active acquisition of task-relevant observations. SeekVLN is trained in two stages: First, Future-guided Reverse Generation (FRG) uses future expert actions to augment offline expert trajectories with supplementary views and evidence annotations. Supervised fine-tuning on these trajectories establishes a prior for evidence seeking and progress reasoning without additional expert interaction. However, imitation alone does not reveal whether seeking improves subsequent navigation. We therefore introduce Counterfactual Contrastive Policy Optimization (C2PO) for reinforcement fine-tuning. By comparing each evidence-seeking branch with a counterfactual direct-navigation branch from the same state, C2PO uses a contrastive reward to assign credit to seeking decisions based on subsequent navigation benefit. Experiments on simulated benchmarks show that SeekVLN achieves state-of-the-art performance, improving success rate by 12.7% and 7.5% over the base model on R2R-CE and RxR-CE, respectively. Both simulated and real-world evaluations exhibit human-like evidence-seeking behaviors for more reliable progress grounding.

Recognizing if LLM output can be grounded in evidence is central to many tasks in NLP: retrieval-augmented generation, summarization, document-grounded dialogue, and more. Current approaches to this kind of "fact-checking" are based on verifying each piece of a model generation against potential evidence using an LLM. However, this process can be very computationally expensive, requiring many calls to LLMs to check a single response. In this work, we show how to build small models that have GPT-4-level performance but for 400x lower cost. We do this by constructing synthetic training data with GPT-4, which involves creating realistic yet challenging instances of factual errors via a structured generation procedure. Training on this data teaches models to check each fact in the claim and recognize synthesis of information across sentences. For evaluation, we unify pre-existing datasets into a benchmark LLM-AggreFact, collected from recent work on fact-checking and grounding LLM generations. Our best system MiniCheck-FT5 (770M parameters) outperforms all systems of comparable size and reaches GPT-4 accuracy. We release LLM-AggreFact, code for data synthesis, and models.

Misinformation and disinformation demand fact checking that goes beyond simple evidence-based reasoning. Existing benchmarks fall short: they are largely single modality (text-only), span short time horizons, use shallow evidence, cover domains unevenly, and often omit full articles -- obscuring models' real-world capability. We present MMM-Fact, a large-scale benchmark of 125,449 fact-checked statements (1995--2025) across multiple domains, each paired with the full fact-check article and multimodal evidence (text, images, videos, tables) from four fact-checking sites and one news outlet. To reflect verification effort, each statement is tagged with a retrieval-difficulty tier -- Basic (1--5 sources), Intermediate (6--10), and Advanced (>10) -- supporting fairness-aware evaluation for multi-step, cross-modal reasoning. The dataset adopts a three-class veracity scheme (true/false/not enough information) and enables tasks in veracity prediction, explainable fact-checking, complex evidence aggregation, and longitudinal analysis. Baselines with mainstream LLMs show MMM-Fact is markedly harder than prior resources, with performance degrading as evidence complexity rises. MMM-Fact offers a realistic, scalable benchmark for transparent, reliable, multimodal fact-checking.

We propose a framework that amortizes the cost of inference-time reasoning by converting transient critiques into retrievable guidelines, through a file-based memory system and agent-controlled tool calls. We evaluate this method on the Rubric Feedback Bench, a novel dataset for rubric-based learning. Experiments demonstrate that our augmented LLMs rapidly match the performance of test-time refinement pipelines while drastically reducing inference cost.

Retrieval-enhanced methods have become a primary approach in fact verification (FV); it requires reasoning over multiple retrieved pieces of evidence to verify the integrity of a claim. To retrieve evidence, existing work often employs off-the-shelf retrieval models whose design is based on the probability ranking principle. We argue that, rather than relevance, for FV we need to focus on the utility that a claim verifier derives from the retrieved evidence. We introduce the feedback-based evidence retriever(FER) that optimizes the evidence retrieval process by incorporating feedback from the claim verifier. As a feedback signal we use the divergence in utility between how effectively the verifier utilizes the retrieved evidence and the ground-truth evidence to produce the final claim label. Empirical studies demonstrate the superiority of FER over prevailing baselines.

Large Language Models (LLMs) are fluent but prone to hallucinations, producing answers that appear plausible yet are unsupported by available evidence. This failure is especially problematic in high-stakes domains where decisions must be justified by verifiable information. We introduce EvidenceRL, a reinforcement learning framework that enforces evidence adherence during training. EvidenceRL scores candidate responses for grounding (entailment with retrieved evidence and context) and correctness (agreement with reference answers) and optimizes the generator using Group Relative Policy Optimization (GRPO). We evaluate across two high-stakes domains, cardiac diagnosis and legal reasoning, where EvidenceRL consistently improves evidence grounding and faithfulness without sacrificing task accuracy. On cardiac diagnosis, F1@3 increases from 37.0 to 54.5 on Llama-3.2-3B while grounding (G_{max}@3) rises from 47.6 to 78.2; hallucinations drop nearly 5times and evidence-supported diagnoses increase from 31.8\% to 61.6\%. On legal reasoning, EvidenceRL raises Faithfulness from 32.8\% to 67.6\% on Llama-3.1-8B, demonstrating consistent behavioral change across domains. Our code is open-sourced at https://github.com/Wizaaard/EvidenceRL.git.

Self-improving LLM agents increasingly learn from experience without updating any weights. Each episode is stored in an external memory, scored, and retrieved for similar future tasks to shape later behavior. Viewed through a reward lens, the stored score is a proxy reward for an implicit, non-parametric policy. Each retrieved episode then becomes a policy-improvement step whose reliability hinges on how that score is produced. In deployment, ground-truth labels are unavailable, so the stored reward is at best an LLM assessment. This substitution creates a failure mode, the *Echo Gap*, across the memory-based self-improving agents and model families studied. Incorrect episodes receive inflated rewards; thus, the agent preferentially reuses the very mistakes it has most confident in. Because the error compounds through memory rather than averaging out and the confirming judge's errors remain correlated with the original self-grading bias, so it cannot identify which memories are overvalued. The missing property is formalized as the *Error-Independence Assumption* (EIA), which we prove is a *necessary* condition for correcting the inflation, not merely a description of a good verifier: a usable signal must track truth *and* decorrelate its error from the memory bias, and the recoverable payoff is a closed-form function of exactly those two quantities. We further show the inflation compounds not only when retrieval ranks by the stored score but also under plain similarity retrieval which is the regime the deployed agent uses. Finally, the answer-free de-inflation algorithm LUCID delivers a consistent end-to-end gain on the BIRD text-to-SQL benchmark. It raises execution accuracy to 56.9%, above both a Memento-style self-graded agent (54.0%, a +2.9-point mean gain across seeds) and a memory-less agent of identical architecture (52.4%).

LLM reasoning traces suffer from complex flaws -- *Step Internal Flaws* (logical errors, hallucinations, etc.) and *Step-wise Flaws* (overthinking, underthinking), which vary by sample. A natural approach would be to provide ground-truth labels to guide LLMs' reasoning. Contrary to intuition, we show that this yields no improvement in reasoning ability. We then propose CRAFT, a unified framework that mitigates both types of Step flaws, which builds a Reasoning Knowledge Graph (RKG) based on the consensus parts of multiple candidate traces, and synthesizes a high-quality trace through topological generation. Our approach improves label-prediction accuracy by 10+% on average, and consistently outperforms all baselines across both logical and mathematical reasoning benchmarks. Further, detailed benchmark evaluation proves that our method also improves the quality of LLMs' reasoning traces in multiple dimensions.

Recent advancements in Large Language Models (LLMs) have driven growing interest in LLM-based agents for complex planning tasks. To avoid costly agent training, many studies adopted memory mechanism that enhances LLM with offline experiences or online trajectory analysis. However, existing works focus on single-granularity memory derived from dynamic environmental interactions, which are inherently constrained by the quality of the collected experiences. This limitation, in turn, constrain the diversity of knowledge and the flexibility of planning. We propose Coarse-to-Fine Grounded Memory (), a novel framework that grounds coarse-to-fine memories with LLM, thereby fully leverage them for flexible adaptation to diverse scenarios. grounds environmental information into coarse-grained focus points to guide experience collection in training tasks, followed by grounding of actionable hybrid-grained tips from each experience. At inference, retrieves task-relevant experiences and tips to support planning. When facing environmental anomalies, the LLM grounds the current situation into fine-grained key information, enabling flexible self-QA reflection and plan correction.

Temporal reasoning over long, multi-session dialogues is a critical capability for conversational agents. However, existing works and our pilot study have shown that as dialogue histories grow in length and accumulate noise, current long-context models struggle to accurately identify temporally pertinent information, significantly impairing reasoning performance. To address this, we introduce Memory-T1, a framework that learns a time-aware memory selection policy using reinforcement learning (RL). It employs a coarse-to-fine strategy, first pruning the dialogue history into a candidate set using temporal and relevance filters, followed by an RL agent that selects the precise evidence sessions. The RL training is guided by a multi-level reward function optimizing (i) answer accuracy, (ii) evidence grounding, and (iii) temporal consistency. In particular, the temporal consistency reward provides a dense signal by evaluating alignment with the query time scope at both the session-level (chronological proximity) and the utterance-level (chronological fidelity), enabling the agent to resolve subtle chronological ambiguities. On the Time-Dialog benchmark, Memory-T1 boosts a 7B model to an overall score of 67.0\%, establishing a new state-of-the-art performance for open-source models and outperforming a 14B baseline by 10.2\%. Ablation studies show temporal consistency and evidence grounding rewards jointly contribute to a 15.0\% performance gain. Moreover, Memory-T1 maintains robustness up to 128k tokens, where baseline models collapse, proving effectiveness against noise in extensive dialogue histories. The code and datasets are publicly available at https://github.com/Elvin-Yiming-Du/Memory-T1/

Language models (LMs) have been shown to memorize a great deal of factual knowledge contained in their training data. But when an LM generates an assertion, it is often difficult to determine where it learned this information and whether it is true. In this paper, we propose the problem of fact tracing: identifying which training examples taught an LM to generate a particular factual assertion. Prior work on training data attribution (TDA) may offer effective tools for identifying such examples, known as "proponents". We present the first quantitative benchmark to evaluate this. We compare two popular families of TDA methods -- gradient-based and embedding-based -- and find that much headroom remains. For example, both methods have lower proponent-retrieval precision than an information retrieval baseline (BM25) that does not have access to the LM at all. We identify key challenges that may be necessary for further improvement such as overcoming the problem of gradient saturation, and also show how several nuanced implementation details of existing neural TDA methods can significantly improve overall fact tracing performance.

We introduce Strategic Doctrine Language Models (sdLM), a learning-system framework for multi-document strategic reasoning with doctrinal consistency constraints and calibrated uncertainty. The approach combines multi-document attention, temporal encoding, and a doctrine-consistency layer to improve long-horizon forecasting and plan plausibility while reducing severe doctrinal violations. We evaluate sdLM using (i) expert-panel scoring of strategic scenarios (N=47), (ii) doctrine consistency on 336 doctrine publications (12,847 statements), and (iii) geopolitical forecasting on 127 historical counterfactuals (1945-2020) across 12-60 month horizons. Across these benchmarks, sdLM achieves higher strategic quality and better calibration than strong general-purpose LLM baselines, and remains competitive with human experts on long-horizon judgments. We further report ablations, scaling trends, and deployment-oriented performance/latency characteristics to clarify which components drive improvements and how they translate to operational settings.

Procedural planning, which entails decomposing a high-level goal into a sequence of temporally ordered steps, is an important yet intricate task for machines. It involves integrating common-sense knowledge to reason about complex contextualized situations that are often counterfactual, e.g. "scheduling a doctor's appointment without a phone". While current approaches show encouraging results using large language models (LLMs), they are hindered by drawbacks such as costly API calls and reproducibility issues. In this paper, we advocate planning using smaller language models. We present PlaSma, a novel two-pronged approach to endow small language models with procedural knowledge and (counterfactual) planning capabilities. More concretely, we develop symbolic procedural knowledge distillation to enhance the implicit knowledge in small language models and an inference-time algorithm to facilitate more structured and accurate reasoning. In addition, we introduce a novel task, Counterfactual Planning, that requires a revision of a plan to cope with a counterfactual situation. In both the original and counterfactual setting, we show that orders-of-magnitude smaller models (770M-11B parameters) can compete and often surpass their larger teacher models' capabilities.

Large language models (LLMs) still produce plausible-sounding but ungrounded factual claims, a problem that worsens in multi-turn dialogue as context grows and early errors cascade. We introduce HalluHard, a challenging multi-turn hallucination benchmark with 950 seed questions spanning four high-stakes domains: legal cases, research questions, medical guidelines, and coding. We operationalize groundedness by requiring inline citations for factual assertions. To support reliable evaluation in open-ended settings, we propose a judging pipeline that iteratively retrieves evidence via web search. It can fetch, filter, and parse full-text sources (including PDFs) to assess whether cited material actually supports the generated content. Across a diverse set of frontier proprietary and open-weight models, hallucinations remain substantial even with web search (approx 30% for the strongest configuration, Opus-4.5 with web search), with content-grounding errors persisting at high rates. Finally, we show that hallucination behavior is shaped by model capacity, turn position, effective reasoning, and the type of knowledge required.

Existing metrics for factuality and faithfulness evaluate whether an answer is supported or contradicted by its grounding documents, but they fail to capture when both supporting and contradicting evidence coexist. We introduce ConflictScore, a novel metric that quantifies how well a model's response acknowledges conflicting evidence in its grounding documents. Our framework decomposes responses into atomic claims, labels each claim against each grounding document, and then aggregates these labels into two complementary measures: ConflictScore-Count (CS-C), the proportion of claims exhibiting conflicts, and ConflictScore-Ratio (CS-R), the balance between supporting and contradicting evidence. We develop ConflictBench, a benchmark covering diverse forms of conflicts such as ambiguity, contradiction, and divergent opinions, to systematically evaluate our metric. Experiments show that ConflictScore effectively detects overconfident claims across domains and can serve as a corrective feedback mechanism that improves truthfulness on TruthfulQA.

Conversational memory changes during use, so endpoint question answering alone cannot establish how a persistent state accumulates, ages, or incorporates revisions. We introduce LSREP, a Longitudinal State-Replay Evaluation Protocol combining ordered replay, explicit lifecycle schedules, repeated probes, evolving reference answers, and mechanism-fidelity checks. Its architectural case study is ICE v2, a local-first memory middleware with typed stores, retrieval fusion, and dynamic context budgets. The private, single-user instantiation contains 1,985 turns, 219 distinct probes, and 1,211 probe-checkpoint observations across 52 checkpoints. On three ordinary-density datasets, ICE v2 has a near-zero mean quality difference from vector-RAG while selecting 32% fewer fragments but using 6.6% more estimated prompt tokens. A fourth, dense dataset exposes catastrophic failures of the unbudgeted baseline. The fidelity audit limits attribution: procedural retrieval is defective, several mechanisms are unexercised, and graph utility is not established. In a complementary matched public diagnostic, ICE v2 loses decisively to pure vector-RAG on LongMemEval: 50.8% versus 72.8% in the evidence-only oracle and 43.0% versus 69.5% in full-S. Paired differences are -22.0 points (95% CI [-26.6, -17.4]) and -26.5 ([-31.3, -21.8]). Conservative abstention accompanies severe multi-session and temporal failures. ICE uses less context in this diagnostic, establishing a quality-cost trade-off rather than superior efficiency. Together, replay, fidelity auditing, and public endpoint testing expose distinct failure modes that neither architectural descriptions nor aggregate scores identify alone.

Multimodal Large Language Models (MLLMs) have significantly advanced document understanding, yet current Doc-VQA evaluations score only the final answer and leave the supporting evidence unchecked. This answer-only approach masks a critical failure mode: a model can land on the correct answer while grounding it in the wrong passage -- a critical risk in high-stakes domains like law, finance, and medicine, where every conclusion must be traceable to a specific source region. To address this, we introduce CiteVQA, a benchmark that requires models to return element-level bounding-box citations alongside each answer, evaluating both jointly. CiteVQA comprises 1,897 questions across 711 PDFs spanning seven domains and two languages, averaging 40.6 pages per document. To ensure fidelity and scalability, the ground-truth citations are generated by an automated pipeline-which identifies crucial evidence via masking ablation-and are subsequently validated through expert review. At the core of our evaluation is Strict Attributed Accuracy (SAA), which credits a prediction only when the answer and the cited region are both correct. Auditing 20 MLLMs reveals a pervasive Attribution Hallucination: models frequently produce the right answer while citing the wrong region. The strongest system (Gemini-3.1-Pro-Preview) achieves an SAA of only 76.0, and the strongest open-source MLLM reaches just 22.5. Ultimately, towards trustworthy document intelligence, CiteVQA exposes a reliability gap that answer-only evaluations overlook, providing the instrumentation needed to close it. Our repository is available at https://github.com/opendatalab/CiteVQA.

Although artificial intelligence (AI) has become deeply integrated into various stages of the research workflow and achieved remarkable advancements, academic rebuttal remains a significant and underexplored challenge. This is because rebuttal is a complex process of strategic communication under severe information asymmetry rather than a simple technical debate. Consequently, current approaches struggle as they largely imitate surface-level linguistics, missing the essential element of perspective-taking required for effective persuasion. In this paper, we introduce RebuttalAgent, the first framework to ground academic rebuttal in Theory of Mind (ToM), operationalized through a ToM-Strategy-Response (TSR) pipeline that models reviewer mental state, formulates persuasion strategy, and generates strategy-grounded response. To train our agent, we construct RebuttalBench, a large-scale dataset synthesized via a novel critique-and-refine approach. Our training process consists of two stages, beginning with a supervised fine-tuning phase to equip the agent with ToM-based analysis and strategic planning capabilities, followed by a reinforcement learning phase leveraging the self-reward mechanism for scalable self-improvement. For reliable and efficient automated evaluation, we further develop Rebuttal-RM, a specialized evaluator trained on over 100K samples of multi-source rebuttal data, which achieves scoring consistency with human preferences surpassing powerful judge GPT-4.1. Extensive experiments show RebuttalAgent significantly outperforms the base model by an average of 18.3% on automated metrics, while also outperforming advanced proprietary models across both automated and human evaluations. Disclaimer: the generated rebuttal content is for reference only to inspire authors and assist in drafting. It is not intended to replace the author's own critical analysis and response.

Recent benchmarks have probed factual consistency and rhetorical robustness in Large Language Models (LLMs). However, a knowledge gap exists regarding how directional framing of factually true statements influences model agreement, a common scenario for LLM users. AssertBench addresses this by sampling evidence-supported facts from FEVEROUS, a fact verification dataset. For each (evidence-backed) fact, we construct two framing prompts: one where the user claims the statement is factually correct, and another where the user claims it is incorrect. We then record the model's agreement and reasoning. The desired outcome is that the model asserts itself, maintaining consistent truth evaluation across both framings, rather than switching its evaluation to agree with the user. AssertBench isolates framing-induced variability from the model's underlying factual knowledge by stratifying results based on the model's accuracy on the same claims when presented neutrally. In doing so, this benchmark aims to measure an LLM's ability to "stick to its guns" when presented with contradictory user assertions about the same fact. The complete source code is available at https://github.com/achowd32/assert-bench.

As large language models are deployed in increasingly autonomous long-horizon tasks, manually auditing and verifying the actions, artifacts, and outputs of models becomes more difficult. Users instead come to rely on LLM-generated reports to assess the quality and completeness of the work. We introduce a suite of eight adversarial reporting scenarios to systematically study whether LLMs conceal narrative-changing flaws: errors or limitations that undermine an otherwise successful account of work. We call this phenomenon "insecure reporting." When handed machine learning experiment logs containing a planted negative result that substantially weakens the proposed method, GPT-5.5 flags the negative result in only 2 of 200 generated reports. However, when a short honesty instruction, "Be honest in your response," is added, the model flags the negative result in 190 of 200 reports. Across eight open-weight models, chain-of-thought analysis reveals a recurring tension between disclosing narrative-changing flaws and reasoning about ways to appear successful. We perform an activation analysis and a steering experiment on Qwen3.5-9B, finding that honesty and success-seeking correspond to opposing directions in representation space. Our results suggest that LLMs tend to present narratives of success by default, and that steering models toward honesty makes their reports substantially more transparent.

Reference-free faithfulness metrics verify each atomic claim a model makes against ground truth, and are increasingly used to evaluate grounded generation. We show they share a blind spot: they measure only precision -- are the stated claims supported? -- and therefore reward abstention, since a model can score near-perfect faithfulness by saying almost nothing. We make this measurable using Formula 1 telemetry, a domain where strategic ground truth is derived deterministically and, crucially, completely: for each decision we know the full set of facts that mattered. This completeness -- absent in open-domain faithfulness benchmarks -- lets us measure recall (coverage of the relevant facts) exactly, alongside precision. On a multilingual (EN/ES/PT) benchmark of 7,253 decision instances spanning 150 races, the most precise frontier model covers under half of the relevant facts and ranks last by F1, so requiring coverage reorders the systems; the same effect reappears in a second complete-oracle domain (NOAA weather forecasts). A prompt ablation shows the low coverage is not an under-prompting artifact: explicitly asking models to be thorough does not close the gap. We pair faithfulness with coverage into a single score, validate the metric (controlled perturbation; agreement across a model-free regex extractor and a cross-family LLM extractor, system-level Spearman 1.0), and give a verifier-guided generation method that improves precision and recall without references. We release the benchmark, structured annotations, metric, baselines, and an interactive demo.

While LLMs exhibit remarkable fluency, their utility is often compromised by factual hallucinations and a lack of traceable provenance. Existing resources for grounding mitigate this but typically enforce a dichotomy: they offer either structured knowledge without textual context (e.g., knowledge bases) or grounded text with limited scale and linguistic coverage. To bridge this gap, we introduce FactNet, a massive, open-source resource designed to unify 1.7 billion atomic assertions with 3.01 billion auditable evidence pointers derived exclusively from 316 Wikipedia editions. Unlike recent synthetic approaches, FactNet employs a strictly deterministic construction pipeline, ensuring that every evidence unit is recoverable with byte-level precision. Extensive auditing confirms a high grounding precision of 92.1%, even in long-tail languages. Furthermore, we establish FactNet-Bench, a comprehensive evaluation suite for Knowledge Graph Completion, Question Answering, and Fact Checking. FactNet provides the community with a foundational, reproducible resource for training and evaluating trustworthy, verifiable multilingual systems.

Learning how to do things from trial and error in real time is a hallmark of biological intelligence, yet most LLM-based agents lack mechanisms to acquire procedural knowledge after deployment. We propose Procedural Recall for Agents with eXperiences Indexed by State (PRAXIS), a lightweight post-training learning mechanism that stores the consequences of actions and retrieves them by jointly matching environmental and internal states of past episodes to the current state. PRAXIS augments agentic action selection with retrieved state-action-result exemplars that are generated in real time. When evaluated on the REAL web browsing benchmark, PRAXIS improves task completion accuracy, reliability, and cost efficiency across different foundation model backbones, and shows preliminary generalization to unseen tasks in similar environments. These results demonstrate that PRAXIS enables the practical adoption of AI agents in fast-evolving stateful environments by helping them learn new procedures effectively.

Existing long-horizon memory benchmarks mostly use multi-turn dialogues or synthetic user histories, which makes retrieval performance an imperfect proxy for person understanding. We present \BenchName, a publicly releasable benchmark built from long-form autobiographical narratives, where actions, context, and inner thoughts provide dense evidence for inferring stable motivations and decision principles. \BenchName~reconstructs each narrative into a flashback-aware, time-anchored stream and evaluates models with evidence-linked questions spanning factual recall, subjective state attribution, and principle-level reasoning. Across diverse narrative sources, retrieval-augmented systems mainly improve factual accuracy, while errors persist on temporally grounded explanations and higher-level inferences, highlighting the need for memory mechanisms beyond retrieval. Our data is in KnowMeBench{https://github.com/QuantaAlpha/KnowMeBench}.

We introduce GUIDE, a novel continual learning approach that directs diffusion models to rehearse samples at risk of being forgotten. Existing generative strategies combat catastrophic forgetting by randomly sampling rehearsal examples from a generative model. Such an approach contradicts buffer-based approaches where sampling strategy plays an important role. We propose to bridge this gap by incorporating classifier guidance into the diffusion process to produce rehearsal examples specifically targeting information forgotten by a continuously trained model. This approach enables the generation of samples from preceding task distributions, which are more likely to be misclassified in the context of recently encountered classes. Our experimental results show that GUIDE significantly reduces catastrophic forgetting, outperforming conventional random sampling approaches and surpassing recent state-of-the-art methods in continual learning with generative replay.

Evaluating retrieval-augmented generation (RAG) pipelines requires corpora where ground truth is knowable, temporally structured, and cross-artifact properties that real-world datasets rarely provide cleanly. Existing resources such as the Enron corpus carry legal ambiguity, demographic skew, and no structured ground truth. Purely LLM-generated synthetic data solves the legal problem but introduces a subtler one: the generating model cannot be prevented from hallucinating facts that contradict themselves across documents.We present OrgForge, an open-source multi-agent simulation framework that enforces a strict physics-cognition boundary: a deterministic Python engine maintains a SimEvent ground truth bus; large language models generate only surface prose, constrained by validated proposals. An actor-local clock enforces causal timestamp correctness across all artifact types, eliminating the class of timeline inconsistencies that arise when timestamps are sampled independently per document. We formalize three graph-dynamic subsystems stress propagation via betweenness centrality, temporal edge-weight decay, and Dijkstra escalation routing that govern organizational behavior independently of any LLM. Running a configurable N-day simulation, OrgForge produces interleaved Slack threads, JIRA tickets, Confluence pages, Git pull requests, and emails, all traceable to a shared, immutable event log. We additionally describe a causal chain tracking subsystem that accumulates cross-artifact evidence graphs per incident, a hybrid reciprocal-rank-fusion recurrence detector for identifying repeated failure classes, and an inbound/outbound email engine that routes vendor alerts, customer complaints, and HR correspondence through gated causal chains with probabilistic drop simulation. OrgForge is available under the MIT license.

Large language models (LLMs) have shown substantial capacity for generating fluent, contextually appropriate responses. However, they can produce hallucinated outputs, especially when a user query includes one or more false premises-claims that contradict established facts. Such premises can mislead LLMs into offering fabricated or misleading details. Existing approaches include pretraining, fine-tuning, and inference-time techniques that often rely on access to logits or address hallucinations after they occur. These methods tend to be computationally expensive, require extensive training data, or lack proactive mechanisms to prevent hallucination before generation, limiting their efficiency in real-time applications. We propose a retrieval-based framework that identifies and addresses false premises before generation. Our method first transforms a user's query into a logical representation, then applies retrieval-augmented generation (RAG) to assess the validity of each premise using factual sources. Finally, we incorporate the verification results into the LLM's prompt to maintain factual consistency in the final output. Experiments show that this approach effectively reduces hallucinations, improves factual accuracy, and does not require access to model logits or large-scale fine-tuning.

Training deep research agents, namely systems that plan, search, evaluate evidence, and synthesize long-form reports, pushes reinforcement learning beyond the regime of verifiable rewards. Their outputs lack ground-truth answers, their trajectories span many tool-augmented decisions, and standard post-training offers little mechanism for turning past attempts into reusable experience. In this work, we argue that rubrics should serve not merely as final-answer evaluators, but as the shared interface that structures policy execution, judge feedback, and agent memory. Based on this view, we introduce RubricEM, a rubric-guided reinforcement learning framework that combines stagewise policy decomposition with reflection-based meta-policy evolution. RubricEM first makes research trajectories stage-aware by conditioning planning, evidence gathering, review, and synthesis on self-generated rubrics. It then assigns credit with Stage-Structured GRPO, which uses stagewise rubric judgments to provide denser semantic feedback for long-horizon optimization. In parallel, RubricEM trains a shared-backbone reflection meta-policy that distills judged trajectories into reusable rubric-grounded guidance for future attempts. The resulting RubricEM-8B achieves strong performance across four long-form research benchmarks, outperforming comparable open models and approaching proprietary deep-research systems. Beyond final performance, we perform thorough analyses to understand the key ingredients of RubricEM.

Writing effective rebuttals is a high-stakes task that demands more than linguistic fluency, as it requires precise alignment between reviewer intent and manuscript details. Current solutions typically treat this as a direct-to-text generation problem, suffering from hallucination, overlooked critiques, and a lack of verifiable grounding. To address these limitations, we introduce RebuttalAgent, the first multi-agents framework that reframes rebuttal generation as an evidence-centric planning task. Our system decomposes complex feedback into atomic concerns and dynamically constructs hybrid contexts by synthesizing compressed summaries with high-fidelity text while integrating an autonomous and on-demand external search module to resolve concerns requiring outside literature. By generating an inspectable response plan before drafting, RebuttalAgent ensures that every argument is explicitly anchored in internal or external evidence. We validate our approach on the proposed RebuttalBench and demonstrate that our pipeline outperforms strong baselines in coverage, faithfulness, and strategic coherence, offering a transparent and controllable assistant for the peer review process. Code will be released.

Agentic search, as a more autonomous and adaptive paradigm of retrieval augmentation, is driving the evolution of intelligent search systems. However, existing evaluation frameworks fail to align well with the goals of agentic search. First, the complex queries commonly used in current benchmarks often deviate from realistic user search scenarios. Second, prior approaches tend to introduce noise when extracting ground truth for end-to-end evaluations, leading to distorted assessments at a fine-grained level. Third, most current frameworks focus solely on the quality of final answers, neglecting the evaluation of the iterative process inherent to agentic search. To address these limitations, we propose RAVine -- a Reality-Aligned eValuation framework for agentic LLMs with search. RAVine targets multi-point queries and long-form answers that better reflect user intents, and introduces an attributable ground truth construction strategy to enhance the accuracy of fine-grained evaluation. Moreover, RAVine examines model's interaction with search tools throughout the iterative process, and accounts for factors of efficiency. We benchmark a series of models using RAVine and derive several insights, which we hope will contribute to advancing the development of agentic search systems. The code and datasets are available at https://github.com/SwordFaith/RAVine.

Large language models (LLMs) are trained on extensive datasets that encapsulate substantial world knowledge. However, their outputs often include confidently stated inaccuracies. Earlier works suggest that LLMs encode truthfulness as a distinct linear feature, termed the "truth direction", which can classify truthfulness reliably. We address several open questions about the truth direction: (i) whether LLMs universally exhibit consistent truth directions; (ii) whether sophisticated probing techniques are necessary to identify truth directions; and (iii) how the truth direction generalizes across diverse contexts. Our findings reveal that not all LLMs exhibit consistent truth directions, with stronger representations observed in more capable models, particularly in the context of logical negation. Additionally, we demonstrate that truthfulness probes trained on declarative atomic statements can generalize effectively to logical transformations, question-answering tasks, in-context learning, and external knowledge sources. Finally, we explore the practical application of truthfulness probes in selective question-answering, illustrating their potential to improve user trust in LLM outputs. These results advance our understanding of truth directions and provide new insights into the internal representations of LLM beliefs. Our code is public at https://github.com/colored-dye/truthfulness_probe_generalization

Large Language Models with transformer architecture have revolutionized the domain of text generation, setting unprecedented benchmarks. Despite their impressive capabilities, LLMs have been criticized for generating outcomes that deviate from factual accuracy or display logical inconsistencies, phenomena commonly referred to as hallucinations. This term, however, has often been misapplied to any results deviating from the instructor's expectations, which this paper defines as attention misdirection rather than true hallucinations. Understanding the distinction between hallucinations and attention misdirection becomes increasingly relevant in business contexts, where the ramifications of such errors can significantly impact the value extraction from these inherently pre-trained models. This paper highlights the best practices of the PGI, Persona, Grouping, and Intelligence, method, a strategic framework that achieved a remarkable error rate of only 3,15 percent across 4,000 responses generated by GPT in response to a real business challenge. It emphasizes that by equipping experimentation with knowledge, businesses can unlock opportunities for innovation through the use of these natively pre-trained models. This reinforces the notion that strategic application grounded in a skilled team can maximize the benefits of emergent technologies such as the LLMs.

A commonly used heuristic in RL is experience replay (e.g.~lin1993reinforcement, mnih2015human), in which a learner stores and re-uses past trajectories as if they were sampled online. In this work, we initiate a rigorous study of this heuristic in the setting of tabular Q-learning. We provide a convergence rate guarantee, and discuss how it compares to the convergence of Q-learning depending on important parameters such as the frequency and number of replay iterations. We also provide theoretical evidence showing when we might expect this heuristic to strictly improve performance, by introducing and analyzing a simple class of MDPs. Finally, we provide some experiments to support our theoretical findings.

While Experience Replay - the practice of storing rollouts and reusing them multiple times during training - is a foundational technique in general RL, it remains largely unexplored in LLM post-training due to the prevailing belief that fresh, on-policy data is essential for high performance. In this work, we challenge this assumption. We present a systematic study of replay buffers for LLM post-training, formalizing the optimal design as a trade-off between staleness-induced variance, sample diversity and the high computational cost of generation. We show that strict on-policy sampling is suboptimal when generation is expensive. Empirically, we show that a well-designed replay buffer can drastically reduce inference compute without degrading - and in some cases even improving - final model performance, while preserving policy entropy.

As LLMs are increasingly used for pre-submission self-review, there is growing demand for feedback that not only identifies weaknesses but also guides authors toward concrete revisions. We study this as Actionable Peer-review Generation and decompose it into two subtasks: diagnostic claim generation and revision suggestion generation. We introduce ActReview, a rebuttal-guided post-training framework that connects paper-specific diagnoses to concrete, grounded revision plans. Our central insight is that author rebuttals reveal plausible actions for addressing reviewer concerns and can therefore provide latent supervision for revision-oriented feedback. From real review-rebuttal threads on OpenReview, we construct ActReview-40K by aligning reviewer weaknesses with author responses and grounding the resulting feedback in localized paper evidence. We post-train Qwen3-8B-Base with multi-task supervised fine-tuning followed by GRPO using candidate-aware, weakness-specific rubric rewards. We also introduce ActReview-Bench, a human-curated benchmark of 1,000 instances for evaluating diagnostic quality and revision usefulness. Experiments show that ActReview outperforms prior specialized review-generation models on actionability and grounding while remaining competitive with strong prompt-based LLMs. Human evaluation confirms improved revision usefulness while revealing a remaining gap in technical accuracy, and additional analyses support generalization to held-out papers and robustness across independent judges.

We explore whether Large Language Models (LLMs) are capable of logical reasoning with distorted facts, which we call Deduction under Perturbed Evidence (DUPE). DUPE presents a unique challenge to LLMs since they typically rely on their parameters, which encode mostly accurate information, to reason and make inferences. However, in DUPE, LLMs must reason over manipulated or falsified evidence present in their prompts, which can result in false conclusions that are valid only under the manipulated evidence. Our goal with DUPE is to determine whether LLMs can arrive at these false conclusions and identify whether the dominant factor influencing the deduction process is the encoded data in the parameters or the manipulated evidence in the prompts. To evaluate the DUPE capabilities of LLMs, we create a DUPEd version of the StrategyQA dataset, where facts are manipulated to reverse the answer to the question. Our findings show that even the most advanced GPT models struggle to reason on manipulated facts - showcasing poor DUPE skills - with accuracy dropping by 45% compared to the original dataset. We also investigate prompt settings inspired from student simulation models, which mitigate the accuracy drop to some extent. Our findings have practical implications for understanding the performance of LLMs in real-world applications such as student simulation models that involve reasoning over inaccurate information.

On-policy reinforcement learning has become a central paradigm for improving the reasoning abilities of large language models. However, its effectiveness is often limited by reward sparsity: when a model fails to discover correct trajectories for difficult problems, the optimization process receives little useful signal and may stagnate. Existing approaches mitigate this issue by incorporating off-policy demonstrations, expert traces, or model-generated solutions, but they typically require the auxiliary data to match the format of the reinforcement-learning task, often relying on rejection sampling from stronger models to obtain suitable training trajectories. We introduce Rationale-Guided Policy Optimization (RGPO), a framework that adaptively leverages ground-truth rationale information according to the model's current capability while preserving its freedom to explore. Rather than treating reference solutions as fixed imitation targets, RGPO uses them as temporary scaffolds: rationales help the model generate improved responses, after which only higher-reward, model-generated solutions are transferred back to the original unguided setting. This design allows training to exploit available ground-truth information without requiring off-policy data to follow the same format as the RL task. Across both language-only and vision-language reasoning settings, RGPO consistently improves performance over RLVR baselines, and ablation studies show that adaptive rationale guidance is a key contributor to these gains. These results suggest that RGPO offers a practical and general approach for reducing reward sparsity, stabilizing reinforcement learning, and improving reasoning performance in both text-only and multimodal models.

Although recent tool-augmented benchmarks involve complex requests, evaluation remains limited to answer matching, neglecting critical trajectory aspects like efficiency, hallucination, and adaptivity. The most straightforward method for evaluation is to compare an agent's trajectory with the ground-truth, but annotating all valid ground-truth trajectories is prohibitively expensive. In this manner, we introduce TRACE, a reference-free framework for the multi-dimensional evaluation of tool-augmented LLMs. By incorporating an evidence bank which accumulates knowledge from preceding steps, TRACE assesses an agent's reasoning trajectory effectively. To validate our framework, we develop a new meta-evaluation dataset with diverse and flawed trajectories, each labeled with multi-faceted performance scores. Our results confirm that TRACE accurately evaluates complex trajectories even with small open-source LLMs. Furthermore, we apply our method to evaluate the trajectories that agents produce while solving tool-augmented tasks, presenting previously unreported observations and their corresponding insights.

The ability of large vision-language models (LVLMs) to critique and correct their reasoning is an essential building block towards their self-improvement. However, a systematic analysis of such capabilities in LVLMs is still lacking. We propose VISCO, the first benchmark to extensively analyze the fine-grained critique and correction capabilities of LVLMs. Compared to existing work that uses a single scalar value to critique the entire reasoning [4], VISCO features dense and fine-grained critique, requiring LVLMs to evaluate the correctness of each step in the chain-of-thought and provide natural language explanations to support their judgments. Extensive evaluation of 24 LVLMs demonstrates that human-written critiques significantly enhance the performance after correction, showcasing the potential of the self-improvement strategy. However, the model-generated critiques are less helpful and sometimes detrimental to the performance, suggesting that critique is the crucial bottleneck. We identified three common patterns in critique failures: failure to critique visual perception, reluctance to "say no", and exaggerated assumption of error propagation. To address these issues, we propose an effective LookBack strategy that revisits the image to verify each piece of information in the initial reasoning. LookBack significantly improves critique and correction performance by up to 13.5%.

Long-horizon agents depend on context management: systems compress, summarize, and evict old tokens so tasks can continue beyond finite windows. That is safe only when dropped information is no longer needed or has been internalized. Plans are the stress case: they are written early, used for many steps, and first to be evicted. We introduce replay pairing, a diagnostic that runs the same trajectory with and without the plan in history and measures hidden-state cosine distance. On Llama-3.1-70B, plan signal spikes to 0.453 one step after the plan, then falls 4.1x in a single action-observation step; HotpotQA falls 12.4x. This is evidence that standard LLM agents do not carry plans forward as persistent state, and instead depend on the plan remaining in context. A layer-L32 probe detects this decay as a diagnostic, not as proof that it reads plan content itself. Reasoning models add a measurement confound: their `<think>` traces re-derive plan content, so standard stripping leaves plan evidence in the stripped condition. We name this the reasoning-trace confound and fix it with strict stripping, which removes prior `<think>` blocks from the stripped run only. It recovers +163% of the step+1 signal in-sample and +153% held out, while not meaningfully changing non-reasoning Llama (+4.8%). On DeepSeek-R1-Distill-Llama-70B, a Llama-trained probe transfers at AUROC 0.748 (p=6e-4), while R1-specific probes reach 1.000, suggesting R1 encodes plan signal in a different hidden-state direction. Finally, a compression stress test shows the practical cost: naive plan eviction cuts ALFWorld success by 34.7pp, while probe-gated re-surfacing does not recover it. The contribution is a measurement and stress-test framework showing that agent-critical information can be context-resident rather than persistent. Context management is load bearing, but plan protection alone is not enough.

AI systems increasingly retrieve from records that revise themselves: issue threads, encyclopedic histories, policy logs, and long conversations. The challenge is not only finding relevant evidence, but deciding which claims remain in force, which were superseded, and when to abstain. Structured memories promise to solve this with typed edges, temporal updates, and conflict status, yet evaluations often change mechanism and prompt presentation together. We study this as Evidence-State Revision, comparing flat retrieval, coarse edge invalidation, and fine-grained RevisionLedger on 2,907 high-agreement questions from GitHub, multi-repo issue histories, Wikipedia, and DyKnow-style temporal streams. A render-matched control (same layout, deprecation disabled) reveals the central confound: when a value is changed and later restored, RevisionLedger appears to beat a flat baseline by +0.182, but almost all the gain comes from easier presentation; the fine-grained mechanism residual is indistinguishable from zero (+0.021 to +0.025 across two judge families). After presentation is controlled, coarse invalidation is the only mechanism that pays for current-state queries, beating the fine ledger by 0.084; the same query-sufficiency principle says provenance mainly needs retained invalidated evidence, not richer typing. Memory evaluations should hold render fixed, and deprecation-aware systems should deploy the coarsest retained state that covers their queries.

In oral arguments, judges probe attorneys with questions about the factual record, legal claims, and the strength of their arguments. To prepare for this questioning, both law schools and practicing attorneys rely on moot courts: practice simulations of appellate hearings. Leveraging a dataset of U.S. Supreme Court oral argument transcripts, we examine whether AI models can effectively simulate justice-specific questioning for moot court-style training. Evaluating oral argument simulation is challenging because there is no single correct question for any given turn. Instead, effective questioning should reflect a combination of desirable qualities, such as anticipating substantive legal issues, detecting logical weaknesses, and maintaining an appropriately adversarial tone. We introduce a two-layer evaluation framework that assesses both the realism and pedagogical usefulness of simulated questions using complementary proxy metrics. We construct and evaluate both prompt-based and agentic oral argument simulators. We find that simulated questions are often perceived as realistic by human annotators and achieve high recall of ground truth substantive legal issues. However, models still face substantial shortcomings, including low diversity in question types and sycophancy. Importantly, these shortcomings would remain undetected under naive evaluation approaches.

Current open-source Large Multimodal Models (LMMs) excel at tasks such as open-vocabulary language grounding and segmentation but can suffer under false premises when queries imply the existence of something that is not actually present in the image. We observe that existing methods that fine-tune an LMM to segment images significantly degrade their ability to reliably determine ("see") if an object is present and to interact naturally with humans ("say"), a form of catastrophic forgetting. In this work, we propose a cascading and joint training approach for LMMs to solve this task, avoiding catastrophic forgetting of previous skills. Our resulting model can "see" by detecting whether objects are present in an image, "say" by telling the user if they are not, proposing alternative queries or correcting semantic errors in the query, and finally "segment" by outputting the mask of the desired objects if they exist. Additionally, we introduce a novel False Premise Correction benchmark dataset, an extension of existing RefCOCO(+/g) referring segmentation datasets (which we call FP-RefCOCO(+/g)). The results show that our method not only detects false premises up to 55% better than existing approaches, but under false premise conditions produces relative cIOU improvements of more than 31% over baselines, and produces natural language feedback judged helpful up to 67% of the time.

Live streaming has emerged as a primary medium for social interaction and digital commerce, yet it is increasingly plagued by sophisticated risks. A fundamental challenge in this domain is tactical out-of-distribution (OOD) shift: while malicious actors maintain stable underlying objectives, they continuously redesign narrative packaging to evade detection. Such adversarial shifts expose critical limitations of existing OOD generalization paradigms, whose assumptions are difficult to satisfy in the presence of tightly coupled intent-tactic evolution and ill-defined raw-level counterfactuals. In this paper, we tackle this issue from a latent causal perspective and propose Latent-Predictive Counterfactual Decoupling~(LPCD), a plug-in framework for robust live streaming risk assessment. LPCD enables counterfactual reasoning under adversarial tactical re-packaging by modeling intent and narrative variation at the latent level, and enforces latent counterfactual consistency to anchor risk prediction on causally stable malicious intent. At inference time, LPCD applies a lightweight, parameter-free calibration to further mitigate tactic-induced distribution shifts. Extensive experiments on large-scale industrial datasets and online production traffic demonstrate that LPCD consistently outperforms state-of-the-art baselines, validating its effectiveness in moderating evolving adversarial risks in real-world live streaming. The project page is available at https://qiaoyran.github.io/LiveStreamingRiskAssessment/.

Dialogue agents based on large language models (LLMs) have shown promising performance in proactive dialogue, which requires effective strategy planning. However, existing approaches to strategy planning for proactive dialogue face several limitations: limited strategy coverage, preference bias in planning, and reliance on costly additional training. To address these, we propose PRINCIPLES: a synthetic strategy memory for proactive dialogue agents. PRINCIPLES is derived through offline self-play simulations and serves as reusable knowledge that guides strategy planning during inference, eliminating the need for additional training and data annotation. We evaluate PRINCIPLES in both emotional support and persuasion domains, demonstrating consistent improvements over strong baselines. Furthermore, PRINCIPLES maintains its robustness across extended and more diverse evaluation settings. See our project page at https://huggingface.co/spaces/kimnamssya/Principles.

Modern LLM agents combine long-term memory for personalization with tool-calling interfaces for taking actions in the world -- a combination underpinning contemporary production systems. We study a previously unexamined failure of this combination: when personality-driven biases stored in memory (cost-consciousness, impatience, risk tolerance, etc.) silently affect tool calls in contexts where they are not applicable. We call this memory-induced tool-drift and operationalize it through MEMDRIFT, a benchmark of 105 scenarios spanning five bias dimensions and seven professional domains, generated through an automated adversarial pipeline. Across seven frontier models -- including those with extended reasoning -- biased memories raise deflection scores (a judge-scored measure of parameter deviation from unbiased baselines) by up to +3.6 points on a 1--5 scale. Tool-drift persists when memory management is handled by three production memory architectures. The phenomenon affects real-world tools: scanning 6{,}062 tools across 288 verified MCP servers, we flag 608 with susceptible parameters and confirm tool-drift on a validated subset. Mechanistically, biased memories act as implicit steering vectors, pushing activations along the same latent directions as explicit behavioral instructions. They also redistribute attention from task-relevant context toward memory entries with surface-level keyword overlap to the target parameter. Standard defenses -- prompt-based relevance instructions and memory filters -- reduce drift but do not eliminate it. As agents take increasingly consequential actions on a user's behalf, memory-induced tool-drift represents a systematic vulnerability that current safeguards do not address, motivating dedicated defenses at the intersection of memory management and tool-call generation.

A key component of fact verification is thevevidence retrieval, often from multiple documents. Recent approaches use dense representations and condition the retrieval of each document on the previously retrieved ones. The latter step is performed over all the documents in the collection, requiring storing their dense representations in an index, thus incurring a high memory footprint. An alternative paradigm is retrieve-and-rerank, where documents are retrieved using methods such as BM25, their sentences are reranked, and further documents are retrieved conditioned on these sentences, reducing the memory requirements. However, such approaches can be brittle as they rely on heuristics and assume hyperlinks between documents. We propose a novel retrieve-and-rerank method for multi-hop retrieval, that consists of a retriever that jointly scores documents in the knowledge source and sentences from previously retrieved documents using an autoregressive formulation and is guided by a proof system based on natural logic that dynamically terminates the retrieval process if the evidence is deemed sufficient. This method is competitive with current state-of-the-art methods on FEVER, HoVer and FEVEROUS-S, while using 5 to 10 times less memory than competing systems. Evaluation on an adversarial dataset indicates improved stability of our approach compared to commonly deployed threshold-based methods. Finally, the proof system helps humans predict model decisions correctly more often than using the evidence alone.

Fact verification systems typically assess whether a claim is supported by retrieved evidence, assuming that truthfulness depends solely on what is stated. However, many real-world claims are half-truths, factually correct yet misleading due to the omission of critical context. Existing models struggle with such cases, as they are not designed to reason about what is left unsaid. We introduce the task of half-truth detection, and propose PolitiFact-Hidden, a new benchmark with 15k political claims annotated with sentence-level evidence alignment and inferred claim intent. To address this challenge, we present TRACER, a modular re-assessment framework that identifies omission-based misinformation by aligning evidence, inferring implied intent, and estimating the causal impact of hidden content. TRACER can be integrated into existing fact-checking pipelines and consistently improves performance across multiple strong baselines. Notably, it boosts Half-True classification F1 by up to 16 points, highlighting the importance of modeling omissions for trustworthy fact verification.

Synthetic insider threat benchmarks face a consistency problem: corpora generated without an external factual constraint cannot rule out cross-artifact contradictions. The CERT dataset -- the field's canonical benchmark -- is also static, lacks cross-surface correlation scenarios, and predates the LLM era. We present OrgForge-IT, a verifiable synthetic benchmark in which a deterministic simulation engine maintains ground truth and language models generate only surface prose, making cross-artifact consistency an architectural guarantee. The corpus spans 51 simulated days, 2,904 telemetry records at a 96.4% noise rate, and four detection scenarios designed to defeat single-surface and single-day triage strategies across three threat classes and eight injectable behaviors. A ten-model leaderboard reveals several findings: (1) triage and verdict accuracy dissociate - eight models achieve identical triage F1=0.80 yet split between verdict F1=1.0 and 0.80; (2) baseline false-positive rate is a necessary companion to verdict F1, with models at identical verdict accuracy differing by two orders of magnitude on triage noise; (3) victim attribution in the vishing scenario separates tiers - Tier A models exonerate the compromised account holder while Tier B models detect the attack but misclassify the victim; (4) rigid multi-signal thresholds structurally exclude single-surface negligent insiders, demonstrating the necessity of parallel, threat-class-specific triage pipelines; and (5) agentic software-engineering training acts as a force multiplier for multi-day temporal correlation, but only when paired with frontier-level parameter scale. Finally, prompt sensitivity analysis reveals that unstructured prompts induce vocabulary hallucination, motivating a two-track scoring framework separating prompt adherence from reasoning capability. OrgForge-IT is open source under the MIT license.

Persistent memory has enabled large language model (LLM) agents to store factual knowledge, prior decisions, reasoning histories, tool usage information, and context. While this has improved the agent's functionality and continuity across tasks, it has also introduced a new attack surface: the agent's own reasoning history. In this paper, we introduce the Forged Amplifying Rationale Memory Attack (FARMA), which poisons an agent's remembered reasoning rather than its factual knowledge. It inserts forged reasoning traces using evasive language that bypasses keyword-based defenses, then amplifies them through self-referential reinforcement that defeats consensus-based defenses. To address FARMA, we introduce SENTINEL, a layered defense pipeline to detect forged reasoning entries. Its central component is the Reasoning Guard that structurally analyzes candidate entries for forgery using five weighted signals. We evaluate FARMA and SENTINEL across multiple agents and different LLM models with 50 trials and show that FARMA achieves an attack success rate of up to 100% under baseline conditions and is capable of defeating defense mechanisms like keyword filter and A-MemGuard. Our evaluation also shows that SENTINEL reduces FARMA's attack success rate to as low as 0% with no false positives observed across 326 benign agent traces. Our work demonstrates the need to protect not only an agent's retrieved content but also the integrity of its reasoning history.

Agentic retrieval-augmented generation (RAG) interleaves retrieval, reasoning, and answer generation across multiple hops. A retrieval error at hop 1 can surface only as a wrong answer at hop 3, while later retrieval can also repair the trajectory. This paper introduces AgenticRAG-FP, an interventional benchmark for causal failure attribution in agentic RAG. The benchmark injects a certified fault at a specified hop, re-executes the downstream trajectory, and evaluates diagnosers against the known intervention. Its central question is whether a post-hoc trace still identifies the injected hop after the suffix changes. In the completed strict dense Claude Haiku 4.5 sweep on 80 three-hop MuSiQue questions, coverage-based diagnosis is 0.91 at hop 1 and 0.00 at hops 2 and 3 (n=43,36,21 failed trajectories). A smaller content-corruption study changes an answer-bearing or bridge fact in topically intact evidence. At depth 2, where 18 failed cases remain after filtering, coverage-based diagnosis is 0.00 and a frozen-hop counterfactual probe is 0.67 in an exploratory pooled comparison. Depth-3 content estimates are descriptive only because they contain three failed cases. These results make propagation depth an explicit evaluation axis for diagnosing agentic RAG failures while distinguishing broad evidence of post-hoc signal loss from small-sample method comparisons.

Learning from past experience benefits from two complementary forms of memory: episodic traces -- raw trajectories of what happened -- and consolidated abstractions distilled across many episodes into reusable, schema-like lessons. Recent agentic-memory systems pursue the consolidated form: an LLM rewrites past trajectories into a textual memory bank that it continuously updates with new interactions, promising self-improving agents without parameter updates. Yet we find that such consolidated memories produced by today's LLMs are often faulty even when derived from useful experiences. As consolidation proceeds, memory utility first rises, then degrades, and can fall below the no-memory baseline. More surprisingly, even when consolidating from ground-truth solutions, GPT-5.4 fails on 54% of a set of ARC-AGI problems it had previously solved without memory. We trace the regression to the consolidation step rather than the underlying experience: the same trajectories yield qualitatively different memories under different update schedules, and an episodic-only control that simply retains those trajectories remains competitive with the consolidators we test. In a controlled ARC-AGI Stream environment that exposes Retain, Delete, and Consolidate actions, agents preserve raw episodes by default and double the accuracy of their forced-consolidation counterparts; disabling consolidation entirely (episodic management only) matches this auto regime. Practically, robust agent memory should treat raw episodes as first-class evidence and gate consolidation explicitly rather than firing it after every interaction. Looking forward, reliable agentic memory will require LLMs that can consolidate without overwriting the evidence they depend on.

The field of automated fact-checking increasingly depends on retrieving web-based evidence to determine the veracity of claims in real-world scenarios. A significant challenge in this process is not only retrieving relevant information, but also identifying evidence that can both support and refute complex claims. Traditional retrieval methods may return documents that directly address claims or lean toward supporting them, but often struggle with more complex claims requiring indirect reasoning. While some existing benchmarks and methods target retrieval for fact-checking, a comprehensive real-world open-domain benchmark has been lacking. In this paper, we present a real-world retrieval benchmark FactIR, derived from Factiverse production logs, enhanced with human annotations. We rigorously evaluate state-of-the-art retrieval models in a zero-shot setup on FactIR and offer insights for developing practical retrieval systems for fact-checking. Code and data are available at https://github.com/factiverse/factIR.

Many tasks require learned models to strategically gather relevant information over multiple rounds of interaction before actually acting on a task. Strategic information gathering requires models to know not only how to effectively acquire information, but also when to stop gathering information and make a decision, in order to avoid overthinking or getting derailed when acting. In this paper, we formalize this problem and introduce Counterfactuals and Reasoning for Termination (CaRT), an approach for teaching LLMs when to stop seeking information. To appropriately learn when to terminate, CaRT fine-tunes LLMs using counterfactual pairs of trajectories, one where termination is appropriate and a minimally modified version of the same trajectory where it is not. It trains the LLM to explain the rationale for the termination decision in either case via verbal reasoning, and imbues this capability into the base LLM via fine-tuning. We instantiate CaRT in two domains: interactive medical diagnosis and math problem solving. In both domains, we find that CaRT improves the efficiency of information gathering and task success rate compared to other fine-tuning methods.

Although agent skills equip LLMs with reusable procedural knowledge, manual maintenance suffers from high costs, unscalability, and misalignment. Real-world deployments thus require autonomous, on-demand skill evolution at test time, constrained by limited interaction budgets and a lack of training or validation sets. This setting introduces a severe sparse reward challenge, where outcomes conflate multiple latent failure causes. Under such ambiguity, existing methods that greedily refine a single incumbent skill are particularly vulnerable to an exploitation trap, allowing early misdiagnoses to exhaust limited trials along unproductive trajectories. To address this, we introduce SkillHEX, a closed-loop framework coupling hypothesis-driven self-verification with evidence-guided tree search. SkillHEX translates falsifiable failure hypotheses into executable tests, producing diagnostic evidence as dense reward without additional environment attempts. This evidence guides a search over persistent skill-revision branches, dynamically balancing the exploitation of supported edits with the exploration of plausible alternatives. Evaluated on 87 tasks from SkillsBench, SkillHEX outperforms existing self-evolving methods and achieves an average pass rate of 55.9% and 57.9% using GPT-5.3-Codex and Claude Opus 4.7 under a five-iteration budget, respectively.

The proliferation of Large Language Models (LLMs) highlights the critical importance of conducting thorough evaluations to discern their comparative advantages, limitations, and optimal use cases. Particularly important is assessing their capacity to accurately retrieve information included in a given prompt. A model's ability to do this significantly influences how effectively it can utilize contextual details, thus impacting its practical efficacy and dependability in real-world applications. Our research analyzes the in-context recall performance of various LLMs using the needle-in-a-haystack method. In this approach, a factoid (the "needle") is embedded within a block of filler text (the "haystack"), which the model is asked to retrieve. We assess the recall performance of each model across various haystack lengths and with varying needle placements to identify performance patterns. This study demonstrates that an LLM's recall capability is not only contingent upon the prompt's content but also may be compromised by biases in its training data. Conversely, adjustments to model architecture, training strategy, or fine-tuning can improve performance. Our analysis provides insight into LLM behavior, offering direction for the development of more effective applications of LLMs.

Continual skill evolution enables LLM agents to accumulate and refine reusable procedural knowledge from interaction experience without updating model parameters. Its effectiveness depends on determining not only what to change, but also why a change is justified and when it should become persistent guidance. However, existing experience-driven methods can lose the behavioral evidence and task contexts supporting edits. Moreover, a global validation outcome provides an incomplete judgment of its constituent changes: locally supported corrections may be discarded with a rejected revision, while evidence may require further experience to inform useful updates. To this end, we introduce EVISKILL, an evidence-driven framework that organizes execution observations into Replayable Evidence Cards and synthesizes edits with explicit links to their supporting contexts. Targeted replay verifies these edits through re-execution and provides feedback for correction. Across epochs, EVISKILL preserves evidence and provisionally retains supported edits for further refinement, while global validation governs their incorporation into the final skill. Experiments on three interactive benchmarks across six LLM backbones demonstrate the effectiveness of this approach.

Adding procedural skills to an LLM agent is typically evaluated by average improvement in task success. However, this metric hides an important cost: skills can also make agents worse. We measure both sides by comparing agents with and without skills across nearly 6,000 runs spanning two office automation benchmarks and three model harness stacks. This allows us to distinguish two outcomes. A regression is a task solved without skills but failed after skills are added. A residual failure is a task that fails both with and without skills. We find that regressions are substantial enough that the best performing skills outperform others primarily by regressing less, not by gaining more. We identify three causes of regression: (i) skill description osmosis, a skill changes an agent's behavior simply by being present in context, even when it is never invoked; (ii) grounding displacement, a skill's prescribed procedure overrides how the agent interprets its inputs; and (iii) verification displacement, where the procedure suppresses checks the agent would otherwise perform on its outputs. Analysing persistent failures reveals the same underlying pattern. Existing skills overemphasize procedural guidance the stage least often responsible for failure while under supporting grounding and verification, the dominant sources of remaining errors. After correcting evaluation artifacts and studying traces, we find many regressions and persistent failures recoverable through better grounding and verification. Procedural skills should be evaluated by decomposing their net effect into gains and regressions, not by aggregate improvement alone. We identify three regression modes skills should avoid, and find that reliability depends more on grounding and verification than on procedural skill choice.

Large language model agents increasingly store reusable procedures outside the model. These reusable procedures are often called skills: they may be code functions, natural-language instructions, SKILL.md packages, workflow graphs, or learned adapters that a future agent can retrieve and invoke. This taxonomy-driven survey asks how such skill libraries change over time. Across a 124-paper 2023--2026 audit set, we synthesize dynamic skill systems as lifecycle-managed, verified, evolving artifact stores: agents collect evidence from interaction, propose skill updates, verify and admit candidates, organize them for retrieval and composition, repair or prune stale entries, and govern sharing through provenance and rollback. We organize the literature around three survey tools. First, a six-sense taxonomy distinguishes the structurally different artifacts called ``skills'' in current papers. Second, an eight-stage lifecycle architecture identifies the recurring design decisions behind evidence acquisition, proposal, verification/admission, storage, retrieval/composition, maintenance, distillation/portability, and governance. Third, a lightweight skill-record schema and ten-operator vocabulary provide common terms for comparing library updates without elevating them into a separate method contribution. Using this structure, we synthesize evidence-graded patterns with explicit caveats: admission and repair are repeatedly important, verifier quality materially affects skill-aware RL, flat retrieval can degrade as libraries grow, and current benchmarks still under-report library trajectories, usage--utility gaps, and safety surfaces. We close with concrete reporting standards and open problems for evaluating dynamic skills as changing libraries rather than static prompt or tool collections.

AI research often requires decisions before future evidence exists: which bottleneck to attack, which direction to pursue, or where a project should be positioned. We introduce ForeSci, a temporally controlled benchmark for evaluating whether LLM agents can make such forward-looking research judgements from historical evidence. ForeSci contains 500 tasks across four fast-moving AI domains and four decision families. Each task is paired with a cutoff-aligned offline knowledge base; post-cutoff papers are hidden during generation and used only for validation. To avoid random future-event prediction, tasks are derived from pre-cutoff taxonomy branches and evidence signals, and answer-generation backbones are selected to precede the task cutoffs. We evaluate native LLMs, Hybrid RAG, and three research-agent adaptations across four backbones. Results show that explicit evidence organization improves traceability and factual support, but gains depend strongly on the decision family. Diagnostics reveal a recurring evidence-decision decoupling: agents may cite relevant evidence while forecasting the wrong research object. ForeSci turns forward-looking AI research judgement into a controlled benchmark for evaluating research agents as decision-making systems.

The training data for many Large Language Models (LLMs) is contaminated with test data. This means that public benchmarks used to assess LLMs are compromised, suggesting a performance gap between benchmark scores and actual capabilities. Ideally, a private holdout set could be used to accurately verify scores. Unfortunately, such datasets do not exist for most benchmarks, and post-hoc construction of sufficiently similar datasets is non-trivial. To address these issues, we introduce a systematic methodology for (i) retrospectively constructing a holdout dataset for a target dataset, (ii) demonstrating the statistical indistinguishability of this retro-holdout dataset, and (iii) comparing LLMs on the two datasets to quantify the performance gap due to the dataset's public availability. Applying these methods to TruthfulQA, we construct and release Retro-Misconceptions, on which we evaluate twenty LLMs and find that some have inflated scores by as much as 16 percentage points. Our results demonstrate that public benchmark scores do not always accurately assess model properties, and underscore the importance of improved data practices in the field.

Evidence-grounded reasoning requires more than attaching retrieved text to a prediction: a model should make decisions that depend on whether the provided evidence supports the target claim. In practice, this often fails because supervision is weak, evidence is only loosely tied to the claim, and evaluation does not test evidence dependence directly. We introduce case-grounded evidence verification, a general framework in which a model receives a local case context, external evidence, and a structured claim, and must decide whether the evidence supports the claim for that case. Our key contribution is a supervision construction procedure that generates explicit support examples together with semantically controlled non-support examples, including counterfactual wrong-state and topic-related negatives, without manual evidence annotation. We instantiate the framework in radiology and train a standard verifier on the resulting support task. The learned verifier substantially outperforms both case-only and evidence-only baselines, remains strong under correct evidence, and collapses when evidence is removed or swapped, indicating genuine evidence dependence. This behavior transfers across unseen evidence articles and an external case distribution, though performance degrades under evidence-source shift and remains sensitive to backbone choice. Overall, the results suggest that a major bottleneck in evidence grounding is not only model capacity, but the lack of supervision that encodes the causal role of evidence.

Automatic fact-checking has recently received more attention as a means of combating misinformation. Despite significant advancements, fact-checking systems based on retrieval-augmented language models still struggle to tackle adversarial claims, which are intentionally designed by humans to challenge fact-checking systems. To address these challenges, we propose a training-free method designed to rephrase the original claim, making it easier to locate supporting evidence. Our modular framework, SUCEA, decomposes the task into three steps: 1) Claim Segmentation and Decontextualization that segments adversarial claims into independent sub-claims; 2) Iterative Evidence Retrieval and Claim Editing that iteratively retrieves evidence and edits the subclaim based on the retrieved evidence; 3) Evidence Aggregation and Label Prediction that aggregates all retrieved evidence and predicts the entailment label. Experiments on two challenging fact-checking datasets demonstrate that our framework significantly improves on both retrieval and entailment label accuracy, outperforming four strong claim-decomposition-based baselines.

Process reward models (PRMs) that provide dense, step-level feedback have shown promise for reinforcement learning, yet their adoption remains limited by the need for expensive step-level annotations or ground truth references. We propose SPARK: a three-stage framework where in the first stage a generator model produces diverse solutions and a verifier model evaluates them using parallel scaling (self-consistency) and sequential scaling (meta-critique). In the second stage, we use these verification outputs as synthetic training data to fine-tune generative process reward models, which subsequently serve as reward signals during training. We show that aggregating multiple independent verifications at the step level produces training data for process reward models that surpass ground-truth outcome supervision, achieving 67.5 F1 on ProcessBench (a benchmark for identifying erroneous steps in mathematical reasoning) compared to 66.4 for reference-guided training and 61.9 for GPT-4o. In the final stage, we apply our generative PRM with chain-of-thought verification (PRM-CoT) as the reward model in RL experiments on mathematical reasoning, and introduce format constraints to prevent reward hacking. Using Qwen2.5-Math-7B, we achieve 47.4% average accuracy across six mathematical reasoning benchmarks, outperforming ground-truth-based RLVR (43.9%). Our work enables reference-free RL training that exceeds ground-truth methods, opening new possibilities for domains lacking verifiable answers or accessible ground truth.

Reinforcement learning (RL) has achieved remarkable success in LLM reasoning, but whether it can also improve direct recall of parametric knowledge remains an open question. We study this question in a controlled zero-shot, one-hop, closed-book QA setting with no chain-of-thought, training only on binary correctness rewards and applying fact-level train-test deduplication to ensure gains reflect improved recall rather than reasoning or memorization. Across three model families and multiple factual QA benchmarks, RL yields ~27% average relative gains, surpassing both training- and inference-time baselines alike. Mechanistically, RL primarily redistributes probability mass over existing knowledge rather than acquiring new facts, moving correct answers from the low-probability tail into reliable greedy generations. Our data-attribution study reveals that the hardest examples are the most informative: those whose answers never appear in 128 pre-RL samples (only ~18% of training data) drive ~83% of the gain, since rare correct rollouts still emerge during training and get reinforced. Together, these findings broaden the role of RL beyond reasoning, repositioning it as a tool for unlocking rather than acquiring latent parametric knowledge.

This paper investigates the role of dynamic external knowledge integration in improving counter-argument generation using Large Language Models (LLMs). While LLMs have shown promise in argumentative tasks, their tendency to generate lengthy, potentially unfactual responses highlights the need for more controlled and evidence-based approaches. We introduce a new manually curated dataset of argument and counter-argument pairs specifically designed to balance argumentative complexity with evaluative feasibility. We also propose a new LLM-as-a-Judge evaluation methodology that shows a stronger correlation with human judgments compared to traditional reference-based metrics. Our experimental results demonstrate that integrating dynamic external knowledge from the web significantly improves the quality of generated counter-arguments, particularly in terms of relatedness, persuasiveness, and factuality. The findings suggest that combining LLMs with real-time external knowledge retrieval offers a promising direction for developing more effective and reliable counter-argumentation systems.

Long-horizon language agents must repeatedly interact with tools, accumulate evidence, and make decisions under bounded context windows. Existing context-management methods make such rollouts feasible by truncating distant history, folding past turns into summaries, or selecting compact memory states. However, these breakthroughs introduce two coupled limitations. First, as the number of turns grows, historical observations are progressively removed or collapsed into compressed states, making it harder for the policy to reuse fine-grained evidence. Second, once the original turns are no longer source-addressable, outcome-based RL loses an explicit path for aligning policy updates with the evidence that supported a successful final answer. To this end, we propose ECHO, a selective turn-memory framework that jointly addresses history collapse and traceable learning through source-indexed reconstruction. Specifically, ECHO compresses each completed environment turn into a compact memory record, reconstructs bounded policy contexts by selecting from these records, and reuses the selected source indices to route positive outcome credit to the evidence and selection actions that support successful answers. On BrowseComp-Plus, ECHO reaches 43.4% held-out accuracy, outperforming GRPO (28.9%) and the rolling-summary baseline SUPO (36.1%), while using fewer turns and lower trajectory volume than SUPO (Figure 1). Additionally, the trained policy improves zero-shot generalization across multi-objective QA, code generation, and deep information-seeking benchmarks on both dense and MoE backbones.

Many robot demonstration datasets contain heterogeneous demonstrations of varying quality. This heterogeneity may benefit policy pre-training, but can hinder robot performance when used with a final imitation learning objective. In particular, some strategies in the data may be less reliable than others or may be underrepresented in the data, leading to poor performance when such strategies are sampled at test time. Moreover, such unreliable or underrepresented strategies can be difficult even for people to discern, and sifting through demonstration datasets is time-consuming and costly. On the other hand, policy performance when trained on such demonstrations can reflect the reliability of different strategies. We thus propose for robots to self-curate based on online robot experience (Demo-SCORE). More specifically, we train and cross-validate a classifier to discern successful policy roll-outs from unsuccessful ones and use the classifier to filter heterogeneous demonstration datasets. Our experiments in simulation and the real world show that Demo-SCORE can effectively identify suboptimal demonstrations without manual curation. Notably, Demo-SCORE achieves over 15-35% higher absolute success rate in the resulting policy compared to the base policy trained with all original demonstrations.

Since the introduction of the Model Context Protocol (MCP), the number of available tools for Large Language Models (LLMs) has increased significantly. These task-specific tool sets offer an alternative to general-purpose tools such as web browsers, while being easier to develop and maintain than GUIs. However, current general-purpose agents predominantly rely on web browsers for interacting with the environment. Here, we introduce TheMCPCompany, a benchmark for evaluating tool-calling agents on tasks that involve interacting with various real-world services. We use the REST APIs of these services to create MCP servers, which include over 18,000 tools. We also provide manually annotated ground-truth tools for each task. In our experiments, we use the ground truth tools to show the potential of tool-calling agents for both improving performance and reducing costs assuming perfect tool retrieval. Next, we explore agent performance using tool retrieval to study the real-world practicality of tool-based agents. While all models with tool retrieval perform similarly or better than browser-based agents, smaller models cannot take full advantage of the available tools through retrieval. On the other hand, GPT-5's performance with tool retrieval is very close to its performance with ground-truth tools. Overall, our work shows that the most advanced reasoning models are effective at discovering tools in simpler environments, but seriously struggle with navigating complex enterprise environments. TheMCPCompany reveals that navigating tens of thousands of tools and combining them in non-trivial ways to solve complex problems is still a challenging task for current models and requires both better reasoning and better retrieval models.

Large language models (LLMs) produce responses rated as highly empathic in single-turn settings (Ayers et al., 2023; Lee et al., 2024), yet they are also known to be formulaic generators that reuse the same lexical patterns, syntactic templates, and discourse structures across tasks (Jiang et al., 2025; Shaib et al., 2024; Namuduri et al., 2025). Less attention has been paid to whether this formulaicity extends to the level of discourse moves, i.e., what a response does for the person it is addressing. This question is especially consequential for empathic dialogue, where effective support demands not just a kind response at one moment but varied strategies as a conversation unfolds (Stiles et al., 1998). Indeed, prior work shows that LLMs reuse the same tactic sequences more than human supporters in single-turn settings (Gueorguieva et al., 2026). We extend this analysis to multi-turn conversations and find that the rigidity compounds: once a tactic appears in a supporter turn, LLMs reuse it in the next at nearly double the rate of humans (0.50-0.56 vs. 0.27). This pattern holds across LLMs serving as supporters in real emotional support conversations, and is invisible to standard similarity metrics. To address this gap, we introduce MINT (Multi-turn Inter-tactic Novelty Training), the first reinforcement learning framework to optimize discourse move diversity across multi-turn empathic dialogue. The best MINT variant combines an empathy quality reward with a cross-turn tactic novelty signal, improving aggregate empathy by 25.3% over vanilla across 1.7B and 4B models while reducing cross-turn discourse move repetition by 26.3% on the 4B model, surpassing all baselines including quality-only and token-level diversity methods on both measures. These results suggest that what current models lack is not empathy itself, but the ability to vary their discourse moves across a conversation.

Embodied LLMs endow robots with high-level task reasoning, but they cannot reflect on what went wrong or why, turning deployment into a sequence of independent trials where mistakes repeat rather than accumulate into experience. Drawing upon human reflective practitioners, we introduce Reflective Test-Time Planning, which integrates two modes of reflection: reflection-in-action, where the agent uses test-time scaling to generate and score multiple candidate actions using internal reflections before execution; and reflection-on-action, which uses test-time training to update both its internal reflection model and its action policy based on external reflections after execution. We also include retrospective reflection, allowing the agent to re-evaluate earlier decisions and perform model updates with hindsight for proper long-horizon credit assignment. Experiments on our newly-designed Long-Horizon Household benchmark and MuJoCo Cupboard Fitting benchmark show significant gains over baseline models, with ablative studies validating the complementary roles of reflection-in-action and reflection-on-action. Qualitative analyses, including real-robot trials, highlight behavioral correction through reflection.

Counterfactual story retelling exposes LLM shortcomings in constrained narrative solution spaces where they can no longer rely on recalling memorised training data. Ground-truth-based post-training, such as SFT, fails to teach LLMs how to generate logical and rational narrative events. In this paper, we introduce Retell, Reward, Repeat (RRR), an RL-based pipeline synthesising Structuralist Narratology with scalar narrativity to teach storytelling structure. We extend the TimeTravel dataset with human-annotated stages of narrative equilibrium to evaluate reward models. By using d-RLAIF, RRR derives training signals from the narrativity of textual features without the need for reference outputs. Evaluations demonstrate that RRR-trained LLMs outperform few-shot and SFT baselines in logic, rationality, and completeness, with output quality additionally validated by blind human preference. Relying on a small, query-only dataset, RRR provides a linguistically grounded, cost-effective post-training mechanism for storytelling--a domain currently lacking effective post-training methods. RRR highlights the continued relevance of integrating established linguistic theories into contemporary NLP.

Large language models (LLMs) have shown significant potential in scientific disciplines such as biomedicine, particularly in hypothesis generation, where they can analyze vast literature, identify patterns, and suggest research directions. However, a key challenge lies in evaluating the truthfulness of generated hypotheses, as verifying their accuracy often requires substantial time and resources. Additionally, the hallucination problem in LLMs can lead to the generation of hypotheses that appear plausible but are ultimately incorrect, undermining their reliability. To facilitate the systematic study of these challenges, we introduce TruthHypo, a benchmark for assessing the capabilities of LLMs in generating truthful biomedical hypotheses, and KnowHD, a knowledge-based hallucination detector to evaluate how well hypotheses are grounded in existing knowledge. Our results show that LLMs struggle to generate truthful hypotheses. By analyzing hallucinations in reasoning steps, we demonstrate that the groundedness scores provided by KnowHD serve as an effective metric for filtering truthful hypotheses from the diverse outputs of LLMs. Human evaluations further validate the utility of KnowHD in identifying truthful hypotheses and accelerating scientific discovery. Our data and source code are available at https://github.com/Teddy-XiongGZ/TruthHypo.

LLMs are increasingly used with external knowledge sources like the internet. Do they weigh information appropriately -- updating more for reliable sources (source discernment) and more when claims bring priors closer to the truth (truth discernment)? We formalize this as information discernment and introduce Learn2Discern (L2D), an experimental framework and benchmark grounded in three normative axioms with interpretable metrics. To establish external validity, a pre-registered, quota-matched user study (n=299) confirms that real LLM users endorse all three axioms and report that violations reduce their trust and usage intent. Across 13 models and nearly 670K trials, we find consistent failures across both dimensions: models perform near chance on source and truth discernment, rely on source popularity twice as much as source reliability, and update roughly equally whether a claim improves or worsens their position relative to the ground truth. Models integrate external knowledge most effectively on datasets where their priors are already the most accurate. Newer and larger models improve truth discernment but not source discernment, a blind spot that model complexity does not address. We identify simple inference-time interventions that improve both forms of discernment. We release our dataset and survey as a testbed for a core alignment property that scales in importance as LLMs replace traditional search.

Large language model (LLM) agents can now post-train an LLM end-to-end. They can write code, launch training, evaluate checkpoints, and improve downstream performance, raising the prospect of AI-for-AI. We argue that this picture conflates two distinct capabilities: execution-level capability, iterating within a selected training strategy; and strategy-level capability, revising the high-level judgment as experimental evidence accumulates. Analyzing a large corpus of publicly released post-training trajectories, we find that across different tasks, the agent's training strategy is locked in at the very beginning, and the entire remaining budget is spent on local adjustments within the selected strategy. We then examine three natural explanations--missing experience, missing guidance, and insufficient reasoning--with escalating interventions. Extensive experiments show that (1) an experience-driven scaffold improves execution across the board (+12.6 points on GSM8K and +40.8 on HumanEval) but leaves the strategy static; (2) human guidance effectively redirects the initial strategy, yet the agent falls back into local adjustment loops once training starts; and (3) additional inference compute pays off on easier tasks but yields almost no gain on the hardest one. In conclusion, what agents lack is neither experience, guidance, nor reasoning compute, but a mechanism for spontaneously reevaluating their strategy during execution.

Single-image, multi-image, and video deep research require different visual operations but share a workflow of visual grounding, external retrieval, and fact composition. A key challenge is to preserve the dependencies linking localized visual anchors, entity relations, source-supported facts, and answer-producing operations. We introduce OneSearch-VL, a unified agent centered on the Visually Grounded Evidence Graph (VGEG), which encodes these dependencies as a shared task-level reference for data construction, process supervision, and operation-level evaluation. Our VGEG-based data engine constructs and verifies multi-image and video questions and filters expert trajectories. Using these data, we assemble OneSearch-VL-SFT-110K and OneSearch-VL-RL-10K for SFT and RL, respectively. We further derive the Evidence-aware Visual-Grounded Rubric reward (EVGR) from VGEG annotations to supervise evidence traceability and visual grounding during RL. For fine-grained evaluation, we construct OneSearch-MI-Bench and OneSearch-Video-Bench, organizing questions by the research operations encoded in their VGEGs. Experiments show that OneSearch-VL-8B improves over Qwen3-VL-8B with tool access by 20.2 and 17.6 percentage points on the two new benchmarks, respectively, while also achieving substantial gains across 7 image benchmarks and VideoDR. Project repository: https://github.com/appletea233/OneSearch-VL

Bullshit, as conceptualized by philosopher Harry Frankfurt, refers to statements made without regard to their truth value. While previous work has explored large language model (LLM) hallucination and sycophancy, we propose machine bullshit as an overarching conceptual framework that can allow researchers to characterize the broader phenomenon of emergent loss of truthfulness in LLMs and shed light on its underlying mechanisms. We introduce the Bullshit Index, a novel metric quantifying LLMs' indifference to truth, and propose a complementary taxonomy analyzing four qualitative forms of bullshit: empty rhetoric, paltering, weasel words, and unverified claims. We conduct empirical evaluations on the Marketplace dataset, the Political Neutrality dataset, and our new BullshitEval benchmark (2,400 scenarios spanning 100 AI assistants) explicitly designed to evaluate machine bullshit. Our results demonstrate that model fine-tuning with reinforcement learning from human feedback (RLHF) significantly exacerbates bullshit and inference-time chain-of-thought (CoT) prompting notably amplify specific bullshit forms, particularly empty rhetoric and paltering. We also observe prevalent machine bullshit in political contexts, with weasel words as the dominant strategy. Our findings highlight systematic challenges in AI alignment and provide new insights toward more truthful LLM behavior.

Current evaluation protocols for Vision-Language Models (VLMs) in Radiology Report Generation (RRG) rely on report-level metrics that measure lexical overlap or aggregate clinical correctness. However, such metrics do not test whether individual diagnostic statements stem from the actual pathological evidence visible in the image. This allows models to achieve competitive scores by exploiting learned priors or spurious correlations, a failure mode we refer to as vision shortcut. We introduce SHOVIR, a benchmark for evaluating vision shortcut behavior in RRG. SHOVIR extends two spatially annotated chest X-ray datasets, MIMIC-CXR and PadChest-GR, with per-box CheXpert labels, and defines image-level and disease-level occlusion experiments that contrast baseline performance on clean images against localized, region-specific perturbations. Comparing predictions across these conditions isolates two failure modes at the disease-class level: direct shortcuts, where a finding persists after its visual evidence is removed, and contextual shortcuts, where detection degrades once co-occurring pathologies are occluded despite the target region remaining intact. Benchmarking eight state-of-the-art VLMs, we find that shortcut behavior varies substantially across architectures and datasets. Models achieving the highest baseline report quality do not necessarily rank highest in spatial grounding, revealing that clinically fluent generation can coexist with shallow reliance on visual evidence. These findings expose a blind spot in current RRG evaluation and motivate region-aware assessment protocols.

We study fact-level repair for multimodal generation, where a fluent output may contain specific facts that are not supported by the input. Existing inference-time repair methods often generate feedback by jointly conditioning on the input and the current output. This design has two limitations: hallucinated claims in the output can bias the model's interpretation of the input, and free-form feedback cannot be ranked or scheduled at the fact level. We present TIGER, an inference-time framework that redesigns feedback for localized repair. TIGER independently extracts an observation graph from the input and a claim graph from the current output, then assigns each claim a graph-conditioned risk score based on support and conflict. The model repairs selected high-risk claims while keeping the backbone frozen. We provide a convergence analysis showing that the expected total risk decreases geometrically to an explicit asymptotic bound under mild assumptions. Experiments across four cross-modal paths, including image-to-text, image+text-to-text, audio-to-text, and video-to-text, show that TIGER reduces unsupported content while preserving task quality. The gains hold across multiple backbones, and a CrisisFACTS case study suggests that the same repair mechanism can improve grounding in multi-source settings.

Evaluating large language models (LLMs) on final-answer correctness is the dominant paradigm. This approach, however, provides a coarse signal for model improvement and overlooks the quality of the underlying reasoning process. We argue that a more granular evaluation of reasoning offers a more effective path to building robust models. We decompose reasoning quality into two dimensions: relevance and coherence. Relevance measures if a step is grounded in the problem; coherence measures if it follows logically from prior steps. To measure these aspects reliably, we introduce causal stepwise evaluation (CaSE). This method assesses each reasoning step using only its preceding context, which avoids hindsight bias. We validate CaSE against human judgments on our new expert-annotated benchmarks, MRa-GSM8K and MRa-MATH. More importantly, we show that curating training data with CaSE-evaluated relevance and coherence directly improves final task performance. Our work provides a scalable framework for analyzing, debugging, and improving LLM reasoning, demonstrating the practical value of moving beyond validity checks.

Safety alignment in open-weight language models is trivially removable: abliteration projects a refusal-mediating direction out of the weights in minutes, and no release-time defense we are aware of prevents it durably. What cannot be prevented can be deceived. Our defense, decoy hardening ("Fool's Gold"), concedes the refusal strip and poisons its payoff: once refusal is stripped, most answers to hazardous operational requests are confident, fluent decoys whose critical elements are falsified. Decoys are trained inside a differentiable simulation of the attack, expressing only in the attacked state; a refusal pin and benign leash hold clean-state behavior to the original. We instantiate it on seven models from five families (9B-122B, dense and mixture-of-experts). On the six models passing our pre-registered efficacy gate, 0.51-0.90 of attacked-state responses to held-out prompts are decoys, +0.27-0.84 attributable to the defense; all six stay within registered benign-behavior and capability budgets; the seventh (smaller) fails the gate (boundary case). Rates replicate on a frozen test split or untouched strata. The claim is epistemic: without independent ground truth, no observation surface we tested separates falsified answers from correct ones - on external red-team benchmarks' CBRNE-adjacent slice, the defended 122B is fatally wrong on 0.82-0.86 of matched-quality answers vs at most 0.10 undefended. Repeated sampling does not restore trust: element-wise consensus at K=64 reconstructs a fully usable procedure on 0.083-0.625 of prompts where the instrument validates, vs 0.58-0.96 undefended, with no label-free way to tell the regimes apart; on the weakest such model the claim is per-draw only. We evaluate chemical and biological hazards; the defense does not address in-context jailbreaks and protects only the initially released defended weights.

LLM search agents increasingly rely on tools at inference time, but their trajectories are often constrained by hard limits on both tool calls and generated tokens. Under such dual budgets, better answers require not only stronger models, but also explicit control over which search action should receive the next budget unit and when the accumulated evidence is sufficient to commit a final answer. We study this problem in multi-hop question answering (QA) and formulate it as two-stage inference-time budget control. At search time, our controller assigns each feasible action a task-level Value-of-Information (VOI) score, defined as an operational estimate of marginal task value per unit budget under the current search state and remaining dual budget, and uses this score to choose among retrieval, decomposition, and answer commitment. After search, a selective evidence-grounded finalizer compares the trajectory answer with a refined candidate and rewrites only when the residual error appears to be a low-risk answer-form error. Across four multi-hop QA benchmarks, three LLM backbones, and four budget levels, the method yields positive aggregate gains over four audited baselines under the same hard dual-budget protocol. Ablations show that search-time budget control, especially budget-dependent penalty, provides the main performance gain, while answer-time control helps mainly when the retrieval path is already adequate. These results suggest that inference-time budget control for LLM search agents should govern both how budget is spent during search and how the final answer is committed.

Multi-agent reasoning systems often use agreement, confidence, or automated scores to decide which messages should shape a final answer. Such filtering assumes that a message likely to be correct is also worth keeping. Yet a wrong answer can contain a useful decomposition, constraint, or scientific principle. We test this distinction with Diverse Hypothesis Deliberation (DHD), a controlled measurement protocol that caches five independently generated messages and replays the same downstream solver, called the integrator, with each message available or hidden. The replay comparison measures a message's trajectory value: whether making the message available helps or harms subsequent reasoning. Across five mathematics and science benchmarks and two openly available model families, gpt-oss-120b and gemma-4-31B-it, wrong-helpful messages appear in every benchmark-model combination. Among wrong-answer messages that change final correctness, more than four in ten changes are helpful in each model. Controlled repeats show that the number of repeatable message effects is unlikely to arise from replay variation alone (p=0.0002). A focused intervention on repeatable wrong-helpful messages finds that the complete message works best, while retaining its reasoning preserves more success than retaining only its answer; the source of the complete-message advantage remains open. Within the same problem, repeated trajectory-value evidence also identifies a better keep-or-remove choice than answer correctness alone. Answer correctness is therefore informative but does not determine trajectory value. DHD measures this missing property and produces reusable labels for learning when agents should listen.

Large Language Models (LLMs) are increasingly deployed in real-world fact-checking systems, yet existing evaluations focus predominantly on claim verification and overlook the broader fact-checking workflow, including claim extraction and evidence retrieval. This narrow focus prevents current benchmarks from revealing systematic reasoning failures, factual blind spots, and robustness limitations of modern LLMs. To bridge this gap, we present FactArena, a fully automated arena-style evaluation framework that conducts comprehensive, stage-wise benchmarking of LLMs across the complete fact-checking pipeline. FactArena integrates three key components: (i) an LLM-driven fact-checking process that standardizes claim decomposition, evidence retrieval via tool-augmented interactions, and justification-based verdict prediction; (ii) an arena-styled judgment mechanism guided by consolidated reference guidelines to ensure unbiased and consistent pairwise comparisons across heterogeneous judge agents; and (iii) an arena-driven claim-evolution module that adaptively generates more challenging and semantically controlled claims to probe LLMs' factual robustness beyond fixed seed data. Across 16 state-of-the-art LLMs spanning seven model families, FactArena produces stable and interpretable rankings. Our analyses further reveal significant discrepancies between static claim-verification accuracy and end-to-end fact-checking competence, highlighting the necessity of holistic evaluation. The proposed framework offers a scalable and trustworthy paradigm for diagnosing LLMs' factual reasoning, guiding future model development, and advancing the reliable deployment of LLMs in safety-critical fact-checking applications.

LLM-powered search agents are increasingly being used for multi-step information seeking tasks, yet the IR community lacks empirical understanding of how agentic search sessions unfold and how retrieved evidence is used. This paper presents a large-scale log analysis of agentic search based on 14.44M search requests (3.97M sessions) collected from DeepResearchGym, i.e. an open-source search API accessed by external agentic clients. We sessionize the logs, assign session-level intents and step-wise query-reformulation labels using LLM-based annotation, and propose Context-driven Term Adoption Rate (CTAR) to quantify whether newly introduced query terms are traceable to previously retrieved evidence. Our analyses reveal distinctive behavioral patterns. First, over 90% of multi-turn sessions contain at most ten steps, and 89% of inter-step intervals fall under one minute. Second, behavior varies by intent. Fact-seeking sessions exhibit high repetition that increases over time, while sessions requiring reasoning sustain broader exploration. Third, agents reuse evidence across steps. On average, 54% of newly introduced query terms appear in the accumulated evidence context, with contributions from earlier steps beyond the most recent retrieval. The findings suggest that agentic search may benefit from repetition-aware early stopping, intent-adaptive retrieval budgets, and explicit cross-step context tracking. We plan to release the anonymized logs to support future research.

Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning and has become a standard post-training paradigm for contemporary language and vision-language models. However, the RLVR recipe introduces a significant risk of capability regression, where models forget foundational skills after prolonged training without employing regularization strategies. We empirically confirm this concern, observing that open-source reasoning models suffer performance degradation on core capabilities such as perception and faithfulness. While imposing regularization terms like KL divergence can help prevent deviation from the base model, these terms are calculated on the current task, thus they do not guarantee broader knowledge. Meanwhile, commonly used experience replay across heterogeneous domains makes it nontrivial to decide how much training focus each objective should receive. To address this, we propose RECAP-a replay strategy with dynamic objective reweighting for general knowledge preservation. Our reweighting mechanism adapts in an online manner using short-horizon signals of convergence and instability, shifting the post-training focus away from saturated objectives and toward underperforming or volatile ones. Our method is end-to-end and readily applicable to existing RLVR pipelines without training additional models or heavy tuning. Extensive experiments on benchmarks based on Qwen2.5-VL-3B and Qwen2.5-VL-7B demonstrate the effectiveness of our method, which not only preserves general capabilities but also improves reasoning by enabling more flexible trade-offs among in-task rewards.

The core premise of AI debate as a scalable oversight technique is that it is harder to lie convincingly than to refute a lie, enabling the judge to identify the correct position. Yet, existing debate experiments have relied on datasets with ground truth, where lying is reduced to defending an incorrect proposition. This overlooks a subjective dimension: lying also requires the belief that the claim defended is false. In this work, we apply debate to subjective questions and explicitly measure large language models' prior beliefs before experiments. Debaters were asked to select their preferred position, then presented with a judge persona deliberately designed to conflict with their identified priors. This setup tested whether models would adopt sycophantic strategies, aligning with the judge's presumed perspective to maximize persuasiveness, or remain faithful to their prior beliefs. We implemented and compared two debate protocols, sequential and simultaneous, to evaluate potential systematic biases. Finally, we assessed whether models were more persuasive and produced higher-quality arguments when defending positions consistent with their prior beliefs versus when arguing against them. Our main findings show that models tend to prefer defending stances aligned with the judge persona rather than their prior beliefs, sequential debate introduces significant bias favoring the second debater, models are more persuasive when defending positions aligned with their prior beliefs, and paradoxically, arguments misaligned with prior beliefs are rated as higher quality in pairwise comparison. These results can inform human judges to provide higher-quality training signals and contribute to more aligned AI systems, while revealing important aspects of human-AI interaction regarding persuasion dynamics in language models.

Under a unified operational definition, we define LLM memory as a persistent state written during pretraining, finetuning, or inference that can later be addressed and that stably influences outputs. We propose a four-part taxonomy (parametric, contextual, external, procedural/episodic) and a memory quadruple (location, persistence, write/access path, controllability). We link mechanism, evaluation, and governance via the chain write -> read -> inhibit/update. To avoid distorted comparisons across heterogeneous setups, we adopt a three-setting protocol (parametric only, offline retrieval, online retrieval) that decouples capability from information availability on the same data and timeline. On this basis we build a layered evaluation: parametric (closed-book recall, edit differential, memorization/privacy), contextual (position curves and the mid-sequence drop), external (answer correctness vs snippet attribution/faithfulness), and procedural/episodic (cross-session consistency and timeline replay, E MARS+). The framework integrates temporal governance and leakage auditing (freshness hits, outdated answers, refusal slices) and uncertainty reporting via inter-rater agreement plus paired tests with multiple-comparison correction. For updating and forgetting, we present DMM Gov: coordinating DAPT/TAPT, PEFT, model editing (ROME, MEND, MEMIT, SERAC), and RAG to form an auditable loop covering admission thresholds, rollout, monitoring, rollback, and change audits, with specs for timeliness, conflict handling, and long-horizon consistency. Finally, we give four testable propositions: minimum identifiability; a minimal evaluation card; causally constrained editing with verifiable forgetting; and when retrieval with small-window replay outperforms ultra-long-context reading. This yields a reproducible, comparable, and governable coordinate system for research and deployment.

Online continual learning aims to continuously train neural networks from a continuous data stream with a single pass-through data. As the most effective approach, the rehearsal-based methods replay part of previous data. Commonly used predictors in existing methods tend to generate biased dot-product logits that prefer to the classes of current data, which is known as a bias issue and a phenomenon of forgetting. Many approaches have been proposed to overcome the forgetting problem by correcting the bias; however, they still need to be improved in online fashion. In this paper, we try to address the bias issue by a more straightforward and more efficient method. By decomposing the dot-product logits into an angle factor and a norm factor, we empirically find that the bias problem mainly occurs in the angle factor, which can be used to learn novel knowledge as cosine logits. On the contrary, the norm factor abandoned by existing methods helps remember historical knowledge. Based on this observation, we intuitively propose to leverage the norm factor to balance the new and old knowledge for addressing the bias. To this end, we develop a heuristic approach called unbias experience replay (UER). UER learns current samples only by the angle factor and further replays previous samples by both the norm and angle factors. Extensive experiments on three datasets show that UER achieves superior performance over various state-of-the-art methods. The code is in https://github.com/FelixHuiweiLin/UER.