Daily Papers

Large language models (LLMs) are central to modern natural language processing, delivering exceptional performance in various tasks. However, their intensive computational and memory requirements present challenges, especially for devices with limited DRAM capacity. This paper tackles the challenge of efficiently running LLMs that exceed the available DRAM capacity by storing the model parameters on flash memory but bringing them on demand to DRAM. Our method involves constructing an inference cost model that harmonizes with the flash memory behavior, guiding us to optimize in two critical areas: reducing the volume of data transferred from flash and reading data in larger, more contiguous chunks. Within this flash memory-informed framework, we introduce two principal techniques. First, "windowing'" strategically reduces data transfer by reusing previously activated neurons, and second, "row-column bundling", tailored to the sequential data access strengths of flash memory, increases the size of data chunks read from flash memory. These methods collectively enable running models up to twice the size of the available DRAM, with a 4-5x and 20-25x increase in inference speed compared to naive loading approaches in CPU and GPU, respectively. Our integration of sparsity awareness, context-adaptive loading, and a hardware-oriented design paves the way for effective inference of LLMs on devices with limited memory.

Six-bit quantization (FP6) can effectively reduce the size of large language models (LLMs) and preserve the model quality consistently across varied applications. However, existing systems do not provide Tensor Core support for FP6 quantization and struggle to achieve practical performance improvements during LLM inference. It is challenging to support FP6 quantization on GPUs due to (1) unfriendly memory access of model weights with irregular bit-width and (2) high runtime overhead of weight de-quantization. To address these problems, we propose TC-FPx, the first full-stack GPU kernel design scheme with unified Tensor Core support of float-point weights for various quantization bit-width. We integrate TC-FPx kernel into an existing inference system, providing new end-to-end support (called FP6-LLM) for quantized LLM inference, where better trade-offs between inference cost and model quality are achieved. Experiments show that FP6-LLM enables the inference of LLaMA-70b using only a single GPU, achieving 1.69x-2.65x higher normalized inference throughput than the FP16 baseline. The source code will be publicly available soon.

Large Language Models (LLMs) have achieved great success in solving difficult tasks across many domains, but such success comes with a high computation cost, and inference latency. As developers and third parties customize these models, the need to provide efficient inference has increased. Many efforts have attempted to reduce inference cost through model compression techniques such as pruning and distillation. However, these techniques either require labeled data, or are time-consuming as they require the compressed model to be retrained to regain accuracy. In this paper, we propose a gradient-free structured pruning framework that uses only unlabeled data. An evaluation on the GLUE and SQuAD benchmarks using BERT_{BASE} and DistilBERT illustrates the effectiveness of the proposed approach. By only using the weights of the pre-trained model and unlabeled data, in a matter of a few minutes on a single GPU, up to 40% of the original FLOP count can be reduced with less than a 4% accuracy loss across all tasks considered.

We introduce a new online adaptive filtering method called supervised multi-step adaptive filters (SMS-AF). Our method uses neural networks to control or optimize linear multi-delay or multi-channel frequency-domain filters and can flexibly scale-up performance at the cost of increased compute -- a property rarely addressed in the AF literature, but critical for many applications. To do so, we extend recent work with a set of improvements including feature pruning, a supervised loss, and multiple optimization steps per time-frame. These improvements work in a cohesive manner to unlock scaling. Furthermore, we show how our method relates to Kalman filtering and meta-adaptive filtering, making it seamlessly applicable to a diverse set of AF tasks. We evaluate our method on acoustic echo cancellation (AEC) and multi-channel speech enhancement tasks and compare against several baselines on standard synthetic and real-world datasets. Results show our method performance scales with inference cost and model capacity, yields multi-dB performance gains for both tasks, and is real-time capable on a single CPU core.

We propose a novel approach for generating complex outputs that significantly improves accuracy in text-to-SQL tasks. Our method leverages execution results to select the most semantically consistent query from multiple candidates, enabling smaller, cost-effective models to surpass computationally intensive reasoning methods such as o1, o3-mini, and DeepSeek R1 while reducing inference cost by as much as 30 times. It integrates effortlessly with existing models, offering a practical and scalable pathway to state-of-the-art SQL generation.

Recently, multiple architectures has been proposed to improve the efficiency of the Transformer Language Models through changing the design of the self-attention block to have a linear-cost inference (LCI). A notable approach in this realm is the State-Space Machines (SSMs) architecture, which showed on-par performance on language modeling tasks with the self-attention transformers. However, such an architectural change requires a full pretraining of the weights from scratch, which incurs a huge cost to researchers and practitioners who want to use the new architectures. In the more traditional linear attention works, it has been proposed to approximate full attention with linear attention by swap-and-finetune framework. Motivated by this approach, we propose Cross-Architecture Transfer Learning (XATL), in which the weights of the shared components between LCI and self-attention-based transformers, such as layernorms, MLPs, input/output embeddings, are directly transferred to the new architecture from already pre-trained model parameters. We experimented the efficacy of the method on varying sizes and alternative attention architectures and show that \methodabbr significantly reduces the training time up to 2.5x times and converges to a better minimum with up to 2.6% stronger model on the LM benchmarks within the same compute budget.

Transformer models, despite their impressive performance, often face practical limitations due to their high computational requirements. At the same time, previous studies have revealed significant activation sparsity in these models, indicating the presence of redundant computations. In this paper, we propose Dynamic Sparsified Transformer Inference (DSTI), a method that radically reduces the inference cost of Transformer models by enforcing activation sparsity and subsequently transforming a dense model into its sparse Mixture of Experts (MoE) version. We demonstrate that it is possible to train small gating networks that successfully predict the relative contribution of each expert during inference. Furthermore, we introduce a mechanism that dynamically determines the number of executed experts individually for each token. DSTI can be applied to any Transformer-based architecture and has negligible impact on the accuracy. For the BERT-base classification model, we reduce inference cost by almost 60%.

Network pruning is widely used for reducing the heavy inference cost of deep models in low-resource settings. A typical pruning algorithm is a three-stage pipeline, i.e., training (a large model), pruning and fine-tuning. During pruning, according to a certain criterion, redundant weights are pruned and important weights are kept to best preserve the accuracy. In this work, we make several surprising observations which contradict common beliefs. For all state-of-the-art structured pruning algorithms we examined, fine-tuning a pruned model only gives comparable or worse performance than training that model with randomly initialized weights. For pruning algorithms which assume a predefined target network architecture, one can get rid of the full pipeline and directly train the target network from scratch. Our observations are consistent for multiple network architectures, datasets, and tasks, which imply that: 1) training a large, over-parameterized model is often not necessary to obtain an efficient final model, 2) learned "important" weights of the large model are typically not useful for the small pruned model, 3) the pruned architecture itself, rather than a set of inherited "important" weights, is more crucial to the efficiency in the final model, which suggests that in some cases pruning can be useful as an architecture search paradigm. Our results suggest the need for more careful baseline evaluations in future research on structured pruning methods. We also compare with the "Lottery Ticket Hypothesis" (Frankle & Carbin 2019), and find that with optimal learning rate, the "winning ticket" initialization as used in Frankle & Carbin (2019) does not bring improvement over random initialization.

Modern LLMs scale at test-time, e.g. via repeated sampling, where inference cost grows with model size and the number of samples. This creates a trade-off that pretraining scaling laws, such as Chinchilla, do not address. We present Train-to-Test (T^2) scaling laws that jointly optimize model size, training tokens, and number of inference samples under fixed end-to-end budgets. T^2 modernizes pretraining scaling laws with pass@k modeling used for test-time scaling, then jointly optimizes pretraining and test-time decisions. Forecasts from T^2 are robust over distinct modeling approaches: measuring joint scaling effect on the task loss and modeling impact on task accuracy. Across eight downstream tasks, we find that when accounting for inference cost, optimal pretraining decisions shift radically into the overtraining regime, well-outside of the range of standard pretraining scaling suites. We validate our results by pretraining heavily overtrained models in the optimal region that T^2 scaling forecasts, confirming their substantially stronger performance compared to pretraining scaling alone. Finally, as frontier LLMs are post-trained, we show that our findings survive the post-training stage, making T^2 scaling meaningful in modern deployments.

Key-Value (KV) cache memory and bandwidth increasingly dominate large language model inference cost in long-context and long-generation regimes. Architectures such as multi-head latent attention (MLA) and hybrid sliding-window attention (SWA) can alleviate this bound, but integrating them into existing models remains difficult. Prior methods impose fine-grained structural requirements on both source and target attention modules, which cannot meet the feasible requirement in practical deployment. We present Attention Editing, a practical framework for converting already-trained large language models (LLMs) with new attention architectures without re-pretraining from scratch. Attention editing replaces the original attention with a learnable target module and trains it using progressive distillation, consisting of (1) layer-wise teacher-forced optimization with intermediate activation supervision to prevent cold-start error accumulation, and (2) model-level distillation on next-token distributions, optionally regularized by weak feature matching. We instantiate the framework on two different target--MLA and GateSWA, a gated hybrid SWA design, and apply it to Qwen3-8B and Qwen3-30B-A3B. The resulting models maintain competitive performance while delivering substantial efficiency improvements, demonstrating that large-scale attention conversion is both feasible and robust. Notably, experiments are conducted on an Ascend 910B clusters, offering a practical training case study on domestic hardware.

Reinforcement learning (RL) post-training has recently driven major gains in long chain-of-thought reasoning large language models (LLMs), but the high inference cost of such models motivates distillation into smaller students. Most existing knowledge distillation (KD) methods are designed for supervised fine-tuning (SFT), relying on fixed teacher traces or teacher-student Kullback-Leibler (KL) divergence-based regularization. When combined with RL, these approaches often suffer from distribution mismatch and objective interference: teacher supervision may not align with the student's evolving rollout distribution, and the KL regularizer can compete with reward maximization and require careful loss balancing. To address these issues, we propose RL-aware distillation (RLAD), which performs selective imitation during RL -- guiding the student toward the teacher only when it improves the current policy update. Our core component, Trust Region Ratio Distillation (TRRD), replaces the teacher-student KL regularizer with a PPO/GRPO-style likelihood-ratio objective anchored to a teacher--old-policy mixture, yielding advantage-aware, trust-region-bounded distillation on student rollouts and naturally balancing exploration, exploitation, and imitation. Across diverse logic reasoning and math benchmarks, RLAD consistently outperforms offline distillation, standard GRPO, and KL-based on-policy teacher-student knowledge distillation.

Agent skills can remarkably improve task success rates by using human-written procedural documents, but their quality is difficult to assess without environment-grounded verification. Existing skill generation methods heavily rely on preference logs rather than direct environment interaction, often yielding negligible or even degraded gains. We identify that it is a fundamental timing bottleneck: robust skills should be posterior-based, distilled from empirical environment interaction rather than prior plans. In this study, we introduce the Posterior Distillation Index (PDI), a trajectory-level metric that quantifies how well a distilled skill is grounded in the task-environment evidence. To operationalize PDI, we present SPARK (Structured Pipelines for Autonomous Runnable tasKs and sKill generation) for preserving task execution evidence towards full trajectory-level analysis. SPARK generates environment-verified trajectories used to compute PDI, and it applies PDI as an online diagnostic and intervention signal to ensure posterior skill formation. Across 86 runnable tasks, SPARK-generated skills consistently surpass no-skill baselines and outperform human-written skills on student models (inference cost up to 1,000x cheaper than teacher models). These findings show that PDI-guided distillation produces efficient and transferable skills grounded in the task-environment interaction. We release our code at https://github.com/EtaYang10th/spark-skills .

Since ChatGPT released its API for public use, the number of applications built on top of commercial large language models (LLMs) increase exponentially. One popular usage of such models is leveraging its in-context learning ability and generating responses given user queries leveraging knowledge obtained by retrieval augmentation. One problem of deploying commercial retrieval-augmented LLMs is the cost due to the additionally retrieved context that largely increases the input token size of the LLMs. To mitigate this, we propose a token compression scheme that includes two methods: summarization compression and semantic compression. The first method applies a T5-based model that is fine-tuned by datasets generated using self-instruct containing samples with varying lengths and reduce token size by doing summarization. The second method further compresses the token size by removing words with lower impact on the semantic. In order to adequately evaluate the effectiveness of the proposed methods, we propose and utilize a dataset called Food-Recommendation DB (FRDB) focusing on food recommendation for women around pregnancy period or infants. Our summarization compression can reduce 65% of the retrieval token size with further 0.3% improvement on the accuracy; semantic compression provides a more flexible way to trade-off the token size with performance, for which we can reduce the token size by 20% with only 1.6% of accuracy drop.

Large language models have emerged as a versatile tool but are challenging to apply to tasks lacking large inference budgets and large in-domain training sets. This work formalizes these constraints and distinguishes four important variables: the pretraining budget (for training before the target domain is known), the specialization budget (for training after the target domain is known), the inference budget, and the in-domain training set size. Across these settings, we compare different approaches from the machine learning literature. Limited by inference cost, we find better alternatives to the standard practice of training very large vanilla transformer models. In particular, we show that hyper-networks and mixture of experts have better perplexity for large pretraining budgets, while small models trained on importance sampled datasets are attractive for large specialization budgets.

The widespread adoption of AI systems in the economy hinges on their ability to generate economic value that outweighs their inference costs. Evaluating this tradeoff requires metrics that account for both performance and costs. We propose a framework grounded in production theory for evaluating language models by combining accuracy and inference cost. We introduce "cost-of-pass", the expected monetary cost of generating a correct solution. We then define the "frontier cost-of-pass" as the minimum cost-of-pass achievable across available models or the "human-expert, using the approximate cost of hiring an expert. Our analysis reveals distinct economic insights. First, lightweight models are most cost-effective for basic quantitative tasks, large models for knowledge-intensive ones, and reasoning models for complex quantitative problems, despite higher per-token costs. Second, tracking this frontier cost-of-pass over the past year reveals significant progress, particularly for complex quantitative tasks where the cost has roughly halved every few months. Third, to trace key innovations driving this progress, we examine counterfactual frontiers: estimates of cost-efficiency without specific model classes. We find that innovations in lightweight, large, and reasoning models have been essential for pushing the frontier in basic quantitative, knowledge-intensive, and complex quantitative tasks, respectively. Finally, we assess the cost-reductions afforded by common inference-time techniques like majority voting and self-refinement, finding that their marginal accuracy gains rarely justify their costs. Our findings underscore that complementary model-level innovations are the primary drivers of cost-efficiency, and our economic framework provides a principled tool for measuring this progress and guiding deployment.

Large language models' significant advances in capabilities are accompanied by significant increases in inference costs. Model routing is a simple technique for reducing inference cost, wherein one maintains a pool of candidate LLMs, and learns to route each prompt to the smallest feasible LLM. Existing works focus on learning a router for a fixed pool of LLMs. In this paper, we consider the problem of dynamic routing, where new, previously unobserved LLMs are available at test time. We propose a new approach to this problem that relies on representing each LLM as a feature vector, derived based on predictions on a set of representative prompts. Based on this, we detail two effective strategies, relying on cluster-based routing and a learned cluster map respectively. We prove that these strategies are estimates of a theoretically optimal routing rule, and provide an excess risk bound to quantify their errors. Experiments on a range of public benchmarks show the effectiveness of the proposed strategies in routing amongst more than 30 unseen LLMs.

Large language model (LLM) scaling laws are empirical formulas that estimate changes in model quality as a result of increasing parameter count and training data. However, these formulas, including the popular DeepMind Chinchilla scaling laws, neglect to include the cost of inference. We modify the Chinchilla scaling laws to calculate the optimal LLM parameter count and pre-training data size to train and deploy a model of a given quality and inference demand. We conduct our analysis both in terms of a compute budget and real-world costs and find that LLM researchers expecting reasonably large inference demand (~1B requests) should train models smaller and longer than Chinchilla-optimal.

Video Diffusion Transformer (DiT) models are a dominant approach for high-quality video generation but suffer from high inference cost due to iterative denoising. Existing caching approaches primarily exploit similarity within the diffusion process of a single request to skip redundant denoising steps. In this paper, we introduce Chorus, a caching approach that leverages similarity across requests to accelerate video diffusion model serving. Chorus achieves up to 45\% speedup on industrial 4-step distilled models, where prior intra-request caching approaches are ineffective. Particularly, Chorus employs a three-stage caching strategy along the denoising process. Stage 1 performs full reuse of latent features from similar requests. Stage 2 exploits inter-request caching in specific latent regions during intermediate denoising steps. This stage is combined with Token-Guided Attention Amplification to improve semantic alignment between the generated video and the conditional prompts, thereby extending the applicability of full reuse to later denoising steps.

With the growing deployment of large language models (LLMs), LLM inference cost has become a key challenge. Pruning techniques that introduce sparsity into weight matrices can accelerate inference. However, maintaining model quality typically limits pruning to moderate unstructured sparsity (around 50\%). At these sparsity levels, none of the existing GPU kernels for sparse matrix multiplication (SpMM) can outperform their dense counterparts. This paper proposes an efficient GPU inference method for LLMs with moderate sparsity. We propose a three-layer matrix storage format comprising: (i) a Sparse-TC layer enabling sparse tensor cores to accelerate SpMM; (ii) a Slot-Filling layer using parallel differential distance for matrix compression while supporting low-cost on-chip decoding; (iii) a lightweight Residual Layer ensuring correct SpMM computation. Building on this format, we design a SpMM kernel that jointly utilizes sparse tensor cores and CUDA cores. This design enables an efficient execution pipeline and overlaps on-chip computation with memory access. Evaluations show that our work is the first to outperform dense matrix multiplication on modern GPUs equipped with high-bandwidth memory (HBM). It achieves up to 1.64x kernel-level speedup over SpInfer (EuroSys'25, Best paper) and up to 1.41x end-to-end speedups over FlashLLM (VLDB'24). Our source code: https://github.com/moui0/cudac.

Vision-language models (VLMs) have achieved impressive performance on multimodal reasoning tasks such as visual question answering (VQA), but their inference cost remains a significant challenge due to the large number of vision tokens processed during the prefill stage. Existing pruning methods often rely on directly using the attention patterns or static text prompt guidance, failing to exploit the dynamic internal signals generated during inference. To address these issues, we propose AdaptInfer, a plug-and-play framework for adaptive vision token pruning in VLMs. First, we introduce a fine-grained, dynamic text-guided pruning mechanism that reuses layer-wise text-to-text attention maps to construct soft priors over text-token importance, allowing more informed scoring of vision tokens at each stage. Second, we perform an offline analysis of cross-modal attention shifts and identify consistent inflection locations in inference, which inspire us to propose a more principled and efficient pruning schedule. Our method is lightweight and plug-and-play, also generalizable across multi-modal tasks. Experimental results have verified the effectiveness of the proposed method. For example, it reduces CUDA latency by 61.3\% while maintaining an average accuracy of 92.9\% on vanilla LLaVA-1.5-7B. Under the same token budget, AdaptInfer surpasses SOTA in accuracy.

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for general-purpose robot control through natural language instructions. However, their high inference cost-stemming from large-scale token computation and autoregressive decoding-poses significant challenges for real-time deployment and edge applications. While prior work has primarily focused on architectural optimization, we take a different perspective by identifying a dual form of redundancy in VLA models: (i) high similarity across consecutive action steps, and (ii) substantial redundancy in visual tokens. Motivated by these observations, we propose FlashVLA, the first training-free and plug-and-play acceleration framework that enables action reuse in VLA models. FlashVLA improves inference efficiency through a token-aware action reuse mechanism that avoids redundant decoding across stable action steps, and an information-guided visual token selection strategy that prunes low-contribution tokens. Extensive experiments on the LIBERO benchmark show that FlashVLA reduces FLOPs by 55.7% and latency by 36.0%, with only a 0.7% drop in task success rate. These results demonstrate the effectiveness of FlashVLA in enabling lightweight, low-latency VLA inference without retraining.

Large language models (LLMs) are widely applied in chatbots, code generators, and search engines. Workloads such as chain-of-thought, complex reasoning, and agent services significantly increase the inference cost by invoking the model repeatedly. Optimization methods such as parallelism, compression, and caching have been adopted to reduce costs, but the diverse service requirements make it hard to select the right method. Recently, specialized LLM inference engines have emerged as a key component for integrating the optimization methods into service-oriented infrastructures. However, a systematic study on inference engines is still lacking. This paper provides a comprehensive evaluation of 25 open-source and commercial inference engines. We examine each inference engine in terms of ease-of-use, ease-of-deployment, general-purpose support, scalability, and suitability for throughput- and latency-aware computation. Furthermore, we explore the design goals of each inference engine by investigating the optimization techniques it supports. In addition, we assess the ecosystem maturity of open source inference engines and handle the performance and cost policy of commercial solutions. We outline future research directions that include support for complex LLM-based services, support of various hardware, and enhanced security, offering practical guidance to researchers and developers in selecting and designing optimized LLM inference engines. We also provide a public repository to continually track developments in this fast-evolving field: https://github.com/sihyeong/Awesome-LLM-Inference-Engine

Scaling the number of parameters and the size of training data has proven to be an effective strategy for improving large language model (LLM) performance. Yet, as these models grow increasingly powerful and widely deployed, the cost of inference has become a pressing concern. Despite its importance, the trade-off between model accuracy and inference efficiency remains underexplored. In this work, we examine how key architectural factors, hidden size, the allocation of parameters between MLP and attention (mlp-to-attention ratio), and grouped-query attention (GQA), influence both inference cost and accuracy. We introduce a conditional scaling law that augments the Chinchilla framework with architectural information, along with a search framework for identifying architectures that are simultaneously inference-efficient and accurate. To validate our approach, we train more than 200 models spanning 80M to 3B parameters and 8B to 100B training tokens, and fit the proposed conditional scaling law. Our results show that the conditional scaling law reliably predicts optimal architectural choices and that the resulting models outperform existing open-source baselines. Under the same training budget, optimized architectures achieve up to 2.1% higher accuracy and 42% greater inference throughput compared to LLaMA-3.2.

Large Language Models (LLMs) and other large foundation models have achieved noteworthy success, but their size exacerbates existing resource consumption and latency challenges. In particular, the large-scale deployment of these models is hindered by the significant resource requirements during inference. In this paper, we study two approaches for mitigating these challenges: employing a cache to store previous queries and learning a model multiplexer to choose from an ensemble of models for query processing. Theoretically, we provide an optimal algorithm for jointly optimizing both approaches to reduce the inference cost in both offline and online tabular settings. By combining a caching algorithm, namely Greedy Dual Size with Frequency (GDSF) or Least Expected Cost (LEC), with a model multiplexer, we achieve optimal rates in both offline and online settings. Empirically, simulations show that the combination of our caching and model multiplexing algorithms greatly improves over the baselines, with up to 50times improvement over the baseline when the ratio between the maximum cost and minimum cost is 100. Experiments on real datasets show a 4.3times improvement in FLOPs over the baseline when the ratio for FLOPs is 10, and a 1.8times improvement in latency when the ratio for average latency is 1.85.

Quantization is an effective technique to reduce the deployment cost of large language models (LLMs), and post-training quantization (PTQ) has been widely studied due to its efficiency. However, existing PTQ methods are limited by their inability to fine-tune model parameters and often suffer significant accuracy loss in low-bit scenarios. Quantization-aware training (QAT) provides a more principled solution, but its reliance on backpropagation incurs prohibitive memory costs, limiting its practicality for LLM deployment. To address these challenges, we propose ZeroQAT, a zeroth-order optimization-based QAT framework that supports both weight and activation quantization. ZeroQAT leverages forward-only gradient estimation to eliminate backpropagation, substantially reducing computational and memory overhead while retaining the benefits of end-to-end optimization. We further introduce a lightweight variant of ZeroQAT for quantized fine-tuning, which freezes and pre-quantizes most parameters to further cut memory usage. Experiments show that ZeroQAT consistently outperforms representative PTQ and QAT baselines while requiring significantly less memory. For example, ZeroQAT enables fine-tuning of a 13B model at extremely low bit-widths (e.g., 2-4 bits) on a single 8GB GPU, and even allows fine-tuning a 6.7B model on a OnePlus 12 smartphone, demonstrating its practicality for end-to-end QAT on resource-limited edge devices.

The dominance of proprietary LLMs has led to restricted access and raised information privacy concerns. High-performing open-source alternatives are crucial for information-sensitive and high-volume applications but often lag behind in performance. To address this gap, we propose (1) A untargeted variant of iterative self-critique and self-refinement devoid of external influence. (2) A novel ranking metric - Performance, Refinement, and Inference Cost Score (PeRFICS) - to find the optimal model for a given task considering refined performance and cost. Our experiments show that SoTA open source models of varying sizes from 7B - 65B, on average, improve 8.2% from their baseline performance. Strikingly, even models with extremely small memory footprints, such as Vicuna-7B, show a 11.74% improvement overall and up to a 25.39% improvement in high-creativity, open ended tasks on the Vicuna benchmark. Vicuna-13B takes it a step further and outperforms ChatGPT post-refinement. This work has profound implications for resource-constrained and information-sensitive environments seeking to leverage LLMs without incurring prohibitive costs, compromising on performance and privacy. The domain-agnostic self-refinement process coupled with our novel ranking metric facilitates informed decision-making in model selection, thereby reducing costs and democratizing access to high-performing language models, as evidenced by case studies.

There is a rapidly growing number of large language models (LLMs) that users can query for a fee. We review the cost associated with querying popular LLM APIs, e.g. GPT-4, ChatGPT, J1-Jumbo, and find that these models have heterogeneous pricing structures, with fees that can differ by two orders of magnitude. In particular, using LLMs on large collections of queries and text can be expensive. Motivated by this, we outline and discuss three types of strategies that users can exploit to reduce the inference cost associated with using LLMs: 1) prompt adaptation, 2) LLM approximation, and 3) LLM cascade. As an example, we propose FrugalGPT, a simple yet flexible instantiation of LLM cascade which learns which combinations of LLMs to use for different queries in order to reduce cost and improve accuracy. Our experiments show that FrugalGPT can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost. The ideas and findings presented here lay a foundation for using LLMs sustainably and efficiently.

Grounded claim factuality checking is important for large language model (LLM) applications such as retrieval-augmented generation, as it helps users assess the correctness of generated outputs. Existing metrics using entailment classifiers require dataset-specific threshold tuning, while LLM-based approaches often use direct prompting, which underutilises the reasoning capabilities of LLMs. We address this by formulating grounded claim factuality checking as a true/false reading comprehension task and prompting LLMs with explicit test-taking strategies for efficient reasoning. Our method reduces token usage by over 80% compared to unguided open-ended reasoning, and achieves competitive performance to more expensive alternatives across two factuality benchmarks, setting a new state of the art on one. To further reduce inference cost, we train small language models (SLMs) to replace LLMs in the checking pipeline. Using supervised fine-tuning (SFT) and a self-revision mechanism, the SLMs learn to improve their factuality judgements. Experimental results show that the resulting SLMs perform on par with strong baselines, combining low inference costs with generating supporting rationales to support interpretability. Code and datasets will be released upon acceptance.

The inference cost of Large Language Models (LLMs) has become a critical factor in determining their commercial viability and widespread adoption. This paper introduces a quantitative ``economics of inference'' framework, treating the LLM inference process as a compute-driven intelligent production activity. We analyze its marginal cost, economies of scale, and quality of output under various performance configurations. Based on empirical data from WiNEval-3.0, we construct the first ``LLM Inference Production Frontier,'' revealing three principles: diminishing marginal cost, diminishing returns to scale, and an optimal cost-effectiveness zone. This paper not only provides an economic basis for model deployment decisions but also lays an empirical foundation for the future market-based pricing and optimization of AI inference resources.

Visual tokens dominate inference cost in vision-language models (VLMs), yet many carry redundant information. Existing pruning methods alleviate this but typically rely on attention magnitude or similarity scores. We reformulate visual token pruning as capacity constrained communication: given a fixed budget K, the model must allocate limited bandwidth to maximally preserve visual information. We propose AutoSelect, which attaches a lightweight Scorer and Denoiser to a frozen VLM and trains with only the standard next token prediction loss, without auxiliary objectives or extra annotations. During training, a variance preserving noise gate modulates each token's information flow according to its predicted importance so that gradients propagate through all tokens; a diagonal attention Denoiser then recovers the perturbed representations. At inference, only the Scorer and a hard top-K selection remain, adding negligible latency. On ten VLM benchmarks, AutoSelect retains 96.5% of full model accuracy while accelerating LLM prefill by 2.85x with only 0.69 ms overhead, and transfers to different VLM backbones without architecture-specific tuning. Code is available at https://github.com/MedHK23/AutoSelect.

To reduce the inference cost of large language models, model compression is increasingly used to create smaller scalable models. However, little is known about their robustness to minority subgroups defined by the labels and attributes of a dataset. In this paper, we investigate the effects of 18 different compression methods and settings on the subgroup robustness of BERT language models. We show that worst-group performance does not depend on model size alone, but also on the compression method used. Additionally, we find that model compression does not always worsen the performance on minority subgroups. Altogether, our analysis serves to further research into the subgroup robustness of model compression.

Diffusion Language Models (DLMs) incur high inference cost due to iterative denoising, motivating efficient pruning. Existing pruning heuristics largely inherited from autoregressive (AR) LLMs, typically preserve attention sink tokens because AR sinks serve as stable global anchors. We show that this assumption does not hold for DLMs: the attention-sink position exhibits substantially higher variance over the full generation trajectory (measured by how the dominant sink locations shift across timesteps), indicating that sinks are often transient and less structurally essential than in AR models. Based on this observation, we propose {bf Sink-Aware Pruning}, which automatically identifies and prunes unstable sinks in DLMs (prior studies usually keep sinks for AR LLMs). Without retraining, our method achieves a better quality-efficiency trade-off and outperforms strong prior pruning baselines under matched compute. Our code is available at https://github.com/VILA-Lab/Sink-Aware-Pruning.

Large Reasoning Models (LRMs) have demonstrated remarkable performance on tasks such as mathematics and code generation. Motivated by these strengths, recent work has empirically demonstrated the effectiveness of LRMs as guard models in improving harmful query detection. However, LRMs typically generate long reasoning traces during inference, causing substantial computational overhead. In this paper, we introduce PSRT, a method that replaces the model's reasoning process with a Prefilled Safe Reasoning Trace, thereby significantly reducing the inference cost of LRMs. Concretely, PSRT prefills "safe reasoning virtual tokens" from a constructed dataset and learns over their continuous embeddings. With the aid of indicator tokens, PSRT enables harmful-query detection in a single forward pass while preserving the classification effectiveness of LRMs. We evaluate PSRT on 7 models, 13 datasets, and 8 jailbreak methods. In terms of efficiency, PSRT completely removes the overhead of generating reasoning tokens during inference. In terms of classification performance, PSRT achieves nearly identical accuracy, with only a minor average F1 drop of 0.015 across 7 models and 5 datasets.

Test-time compute scaling has emerged as a new axis along which to improve model accuracy, where additional computation is used at inference time to allow the model to think longer for more challenging problems. One promising approach for test-time compute scaling is search against a process reward model, where a model generates multiple potential candidates at each step of the search, and these partial trajectories are then scored by a separate reward model in order to guide the search process. The diversity of trajectories in the tree search process affects the accuracy of the search, since increasing diversity promotes more exploration. However, this diversity comes at a cost, as divergent trajectories have less KV sharing, which means they consume more memory and slow down the search process. Previous search methods either do not perform sufficient exploration, or else explore diverse trajectories but have high latency. We address this challenge by proposing Efficient Tree Search (ETS), which promotes KV sharing by pruning redundant trajectories while maintaining necessary diverse trajectories. ETS incorporates a linear programming cost model to promote KV cache sharing by penalizing the number of nodes retained, while incorporating a semantic coverage term into the cost model to ensure that we retain trajectories which are semantically different. We demonstrate how ETS can achieve 1.8times reduction in average KV cache size during the search process, leading to 1.4times increased throughput relative to prior state-of-the-art methods, with minimal accuracy degradation and without requiring any custom kernel implementation. Code is available at: https://github.com/SqueezeAILab/ETS.

Multimodal large language models (MLLMs) incur substantial inference cost due to the processing of hundreds of visual tokens per image. Although token pruning has proven effective for accelerating inference, determining when and where to prune remains largely heuristic. Existing approaches typically rely on static, empirically selected layers, which limit interpretability and transferability across models. In this work, we introduce a matrix-entropy perspective and identify an "Entropy Collapse Layer" (ECL), where the information content of visual representations exhibits a sharp and consistent drop, which provides a principled criterion for selecting the pruning stage. Building on this observation, we propose EntropyPrune, a novel matrix-entropy-guided token pruning framework that quantifies the information value of individual visual tokens and prunes redundant ones without relying on attention maps. Moreover, to enable efficient computation, we exploit the spectral equivalence of dual Gram matrices, reducing the complexity of entropy computation and yielding up to a 64x theoretical speedup. Extensive experiments on diverse multimodal benchmarks demonstrate that EntropyPrune consistently outperforms state-of-the-art pruning methods in both accuracy and efficiency. On LLaVA-1.5-7B, our method achieves a 68.2% reduction in FLOPs while preserving 96.0% of the original performance. Furthermore, EntropyPrune generalizes effectively to high-resolution and video-based models, highlighting the strong robustness and scalability in practical MLLM acceleration. The code will be publicly available at https://github.com/YahongWang1/EntropyPrune.

Vision-Language Models (VLMs) such as CLIP are trained on large amounts of image-text pairs, resulting in remarkable generalization across several data distributions. The prohibitively expensive training and data collection/curation costs of these models make them valuable Intellectual Property (IP) for organizations. This motivates a vendor-client paradigm, where a vendor trains a large-scale VLM and grants only input-output access to clients on a pay-per-query basis in a black-box setting. The client aims to minimize inference cost by distilling the VLM to a student model using the limited available task-specific data, and further deploying this student model in the downstream application. While naive distillation largely improves the In-Domain (ID) accuracy of the student, it fails to transfer the superior out-of-distribution (OOD) generalization of the VLM teacher using the limited available labeled images. To mitigate this, we propose Vision-Language to Vision-Align, Distill, Predict (VL2V-ADiP), which first aligns the vision and language modalities of the teacher model with the vision modality of a pre-trained student model, and further distills the aligned VLM embeddings to the student. This maximally retains the pre-trained features of the student, while also incorporating the rich representations of the VLM image encoder and the superior generalization of the text embeddings. The proposed approach achieves state-of-the-art results on the standard Domain Generalization benchmarks in a black-box teacher setting, and also when weights of the VLM are accessible.

Auto-regressive large language models (LLMs) have yielded impressive performance in many real-world tasks. However, the new paradigm of these LLMs also exposes novel threats. In this paper, we explore their vulnerability to inference cost attacks, where a malicious user crafts Engorgio prompts to intentionally increase the computation cost and latency of the inference process. We design Engorgio, a novel methodology, to efficiently generate adversarial Engorgio prompts to affect the target LLM's service availability. Engorgio has the following two technical contributions. (1) We employ a parameterized distribution to track LLMs' prediction trajectory. (2) Targeting the auto-regressive nature of LLMs' inference process, we propose novel loss functions to stably suppress the appearance of the <EOS> token, whose occurrence will interrupt the LLM's generation process. We conduct extensive experiments on 13 open-sourced LLMs with parameters ranging from 125M to 30B. The results show that Engorgio prompts can successfully induce LLMs to generate abnormally long outputs (i.e., roughly 2-13times longer to reach 90%+ of the output length limit) in a white-box scenario and our real-world experiment demonstrates Engergio's threat to LLM service with limited computing resources. The code is accessible at https://github.com/jianshuod/Engorgio-prompt.

Recent advancements in timestep-distilled diffusion models have enabled high-quality image generation that rivals non-distilled multi-step models, but with significantly fewer inference steps. While such models are attractive for applications due to the low inference cost and latency, fine-tuning them with a naive diffusion objective would result in degraded and blurry outputs. An intuitive alternative is to repeat the diffusion distillation process with a fine-tuned teacher model, which produces good results but is cumbersome and computationally intensive; the distillation training usually requires magnitude higher of training compute compared to fine-tuning for specific image styles. In this paper, we present an algorithm named pairwise sample optimization (PSO), which enables the direct fine-tuning of an arbitrary timestep-distilled diffusion model. PSO introduces additional reference images sampled from the current time-step distilled model, and increases the relative likelihood margin between the training images and reference images. This enables the model to retain its few-step generation ability, while allowing for fine-tuning of its output distribution. We also demonstrate that PSO is a generalized formulation which can be flexibly extended to both offline-sampled and online-sampled pairwise data, covering various popular objectives for diffusion model preference optimization. We evaluate PSO in both preference optimization and other fine-tuning tasks, including style transfer and concept customization. We show that PSO can directly adapt distilled models to human-preferred generation with both offline and online-generated pairwise preference image data. PSO also demonstrates effectiveness in style transfer and concept customization by directly tuning timestep-distilled diffusion models.

Recent work has explored generative recommender systems as an alternative to traditional ID-based models, reframing item recommendation as a sequence generation task over discrete item tokens. While promising, such methods often underperform in practice compared to well-tuned ID-based baselines like SASRec. In this paper, we identify two key limitations holding back generative approaches: the lack of collaborative signal in item tokenization, and inefficiencies in the commonly used encoder-decoder architecture. To address these issues, we introduce COSETTE, a contrastive tokenization method that integrates collaborative information directly into the learned item representations, jointly optimizing for both content reconstruction and recommendation relevance. Additionally, we propose MARIUS, a lightweight, audio-inspired generative model that decouples timeline modeling from item decoding. MARIUS reduces inference cost while improving recommendation accuracy. Experiments on standard sequential recommendation benchmarks show that our approach narrows, or even eliminates, the performance gap between generative and modern ID-based models, while retaining the benefits of the generative paradigm.

Deploying pre-trained transformer models like BERT on downstream tasks in resource-constrained scenarios is challenging due to their high inference cost, which grows rapidly with input sequence length. In this work, we propose a constraint-aware and ranking-distilled token pruning method ToP, which selectively removes unnecessary tokens as input sequence passes through layers, allowing the model to improve online inference speed while preserving accuracy. ToP overcomes the limitation of inaccurate token importance ranking in the conventional self-attention mechanism through a ranking-distilled token distillation technique, which distills effective token rankings from the final layer of unpruned models to early layers of pruned models. Then, ToP introduces a coarse-to-fine pruning approach that automatically selects the optimal subset of transformer layers and optimizes token pruning decisions within these layers through improved L_0 regularization. Extensive experiments on GLUE benchmark and SQuAD tasks demonstrate that ToP outperforms state-of-the-art token pruning and model compression methods with improved accuracy and speedups. ToP reduces the average FLOPs of BERT by 8.1x while achieving competitive accuracy on GLUE, and provides a real latency speedup of up to 7.4x on an Intel CPU.

Diffusion models are rapidly redefining 3D anomaly detection in point cloud data. As 3D sensing becomes integral to modern manufacturing, reliable anomaly detection is essential for high-throughput quality assurance and process control. Yet practical deployment on resource-constrained, latency-critical systems remains limited. Existing methods are often computationally prohibitive or unreliable in complex, unmasked regions, and diffusion pipelines are inherently bottlenecked by iterative denoising. In this work, we address this bottleneck by reformulating reconstructionbased anomaly detection through consistency learning, enabling direct prediction of anomaly-free geometry in one or two network evaluations. We further introduce a novel hybrid loss formulation that explicitly enforces reconstruction toward clean data. This design substantially reduces inference cost, achieving up to 80x faster runtime than the current state-of-the-art method, without GPU acceleration, while preserving strong detection performance. It outperforms R3D-AD on Anomaly-ShapeNet with 76.20% I-AUROC and remains competitive on Real3DAD with 72.80% I-AUROC, enabling efficient, low-latency anomaly detection on resource-constrained platforms, including drones, smart industrial cameras, and other edge devices.

Large reasoning models rely on long chain-of-thought to achieve strong performance, but applying such reasoning uniformly incurs high computational cost. Existing efficiency-oriented methods attempt to shorten or mix reasoning strategies, yet often degrade reasoning capability. We identify the root cause as sequence-level coupling between efficiency incentives and correctness optimization, which implicitly penalizes long but correct reasoning trajectories. To address this issue, we propose Adaptive Dual-Process Thinking (ADaPT), a token-level dual-process framework that explicitly decouples efficiency and correctness signals during training. ADaPT introduces a mode-selection token to control fast and slow reasoning, applying efficiency-related rewards exclusively to this token to avoid penalizing correct long reasoning while encouraging efficiency when appropriate. Moreover, ADaPT enables precise and continuous control over the efficiency-performance trade-off at inference time: by adjusting the generation probability of the mode-selection token, a single trained model can smoothly move along the efficiency-performance Pareto frontier. Extensive experiments demonstrate that ADaPT significantly reduces inference cost while maintaining strong reasoning performance across multiple benchmarks.

Auto-regressive Large Language Models (LLMs) achieve strong performance on coding tasks, but incur high memory and inference costs. Diffusion-based language models (d-LLMs) offer bounded inference cost via iterative denoising, but their behavior under post-training quantization (PTQ) has been sparsely explored. We investigate the application and robustness of PTQ techniques, specifically GPTQ and a modified Hessian-Aware Quantization (HAWQ) algorithm, on a diffusion-based coding LLM (CoDA) and observe that these methods applied to CoDA exhibit greater robustness at low bitwidths compared to Qwen3-1.7B, its auto-regressive counterpart, under a standardized evaluation pipeline. We find that in our setup, CoDA exhibits greater robustness at low bitwidths (2-4 bits), with smaller accuracy degradation across HumanEval and MBPP benchmarks. Additionally, mixed-precision configurations derived from HAWQ provide smooth trade-offs across accuracy, latency, and memory. The results suggest that diffusion LLMs may offer advantages for efficient deployment due to more quantization-resilience.

Large reasoning models (LRMs) extend large language models with explicit multi-step reasoning traces, but this capability introduces a new class of prompt-induced inference-time denial-of-service (PI-DoS) attacks that exploit the high computational cost of reasoning. We first formalize inference cost for LRMs and define PI-DoS, then prove that any practical PI-DoS attack should satisfy three properties: (1) a high amplification ratio, where each query induces a disproportionately long reasoning trace relative to its own length; (ii) stealthiness, in which prompts and responses remain on the natural language manifold and evade distribution shift detectors; and (iii) optimizability, in which the attack supports efficient optimization without being slowed by its own success. Under this framework, we present ReasoningBomb, a reinforcement-learning-based PI-DoS framework that is guided by a constant-time surrogate reward and trains a large reasoning-model attacker to generate short natural prompts that drive victim LRMs into pathologically long and often effectively non-terminating reasoning. Across seven open-source models (including LLMs and LRMs) and three commercial LRMs, ReasoningBomb induces 18,759 completion tokens on average and 19,263 reasoning tokens on average across reasoning models. It outperforms the the runner-up baseline by 35% in completion tokens and 38% in reasoning tokens, while inducing 6-7x more tokens than benign queries and achieving 286.7x input-to-output amplification ratio averaged across all samples. Additionally, our method achieves 99.8% bypass rate on input-based detection, 98.7% on output-based detection, and 98.4% against strict dual-stage joint detection.

Transformer-based large Language Models (LLMs) become increasingly important in various domains. However, the quadratic time complexity of attention operation poses a significant challenge for scaling to longer contexts due to the extremely high inference latency and GPU memory consumption for caching key-value (KV) vectors. This paper proposes RetrievalAttention, a training-free approach to accelerate attention computation. To leverage the dynamic sparse property of attention, RetrievalAttention builds approximate nearest neighbor search (ANNS) indexes upon KV vectors in CPU memory and retrieves the most relevant ones via vector search during generation. Due to the out-of-distribution (OOD) between query vectors and key vectors, off-the-shelf ANNS indexes still need to scan O(N) (usually 30% of all keys) data for accurate retrieval, which fails to exploit the high sparsity. RetrievalAttention first identifies the OOD challenge of ANNS-based attention, and addresses it via an attention-aware vector search algorithm that can adapt to queries and only access 1--3% of data, thus achieving a sub-linear time complexity. RetrievalAttention greatly reduces the inference cost of long-context LLM with much lower GPU memory requirements while maintaining the model accuracy. Especially, RetrievalAttention only needs 16GB GPU memory for serving 128K tokens in LLMs with 8B parameters, which is capable of generating one token in 0.188 seconds on a single NVIDIA RTX4090 (24GB).

Scaling model size and training data has led to great advances in the performance of Large Language Models (LLMs). However, the diminishing returns of this approach necessitate alternative methods to improve model capabilities, particularly in tasks requiring advanced reasoning. Large reasoning models, which leverage long chain-of-thoughts, bring unprecedented breakthroughs in problem-solving capabilities but at a substantial deployment cost associated to longer generations. Reducing inference costs is crucial for the economic feasibility, user experience, and environmental sustainability of these models. In this work, we propose to train large reasoning models to reason efficiently. More precisely, we use reinforcement learning (RL) to train reasoning models to dynamically allocate inference-time compute based on task complexity. Our method incentivizes models to minimize unnecessary computational overhead while maintaining accuracy, thereby achieving substantial efficiency gains. It enables the derivation of a family of reasoning models with varying efficiency levels, controlled via a single hyperparameter. Experiments on two open-weight large reasoning models demonstrate significant reductions in inference cost while preserving most of the accuracy.

Implicit Chain-of-Thought (CoT) reduces the inference cost of large language models by internalizing the explicit rationales. However, existing approaches typically lack alignment with explicit rationales and adaptivity to example complexity. In this work, we propose CIRF (\underline{Chain-of-thoughts Into Reusable Functional units}), an implicit CoT framework that performs reasoning as a dynamic sequence of discrete functional tokens. CIRF assigns a functional token to each semantically coherent reasoning unit in explicit CoT traces. The model is then fine-tuned to autoregressively generate functional tokens and their optional results, followed by the final answer. This design aligns latent reasoning with a sequence of functional units, facilitating parallel training, explicit rationale alignment, and adaptive reasoning. Extensive experiments on mathematical, symbolic, and commonsense reasoning benchmarks show that CIRF provides a favorable accuracy-latency trade-off compared with state-of-the-art implicit CoT methods. Further analyses demonstrate that CIRF constructs distinct, interpretable functional tokens, leading to consistent performance improvements.

Latent reasoning models (LRMs) have attracted significant research interest due to their low inference cost (relative to explicit reasoning models) and theoretical ability to explore multiple reasoning paths in parallel. However, these benefits come at the cost of reduced interpretability: LRMs are difficult to monitor because they do not reason in natural language. This paper presents an investigation into LRM interpretability by examining two state-of-the-art LRMs. First, we find that latent reasoning tokens are often unnecessary for LRMs' predictions; on logical reasoning datasets, LRMs can almost always produce the same final answers without using latent reasoning at all. This underutilization of reasoning tokens may partially explain why LRMs do not consistently outperform explicit reasoning methods and raises doubts about the stated role of these tokens in prior work. Second, we demonstrate that when latent reasoning tokens are necessary for performance, we can decode gold reasoning traces up to 65-93% of the time for correctly predicted instances. This suggests LRMs often implement the expected solution rather than an uninterpretable reasoning process. Finally, we present a method to decode a verified natural language reasoning trace from latent tokens without knowing a gold reasoning trace a priori, demonstrating that it is possible to find a verified trace for a majority of correct predictions but only a minority of incorrect predictions. Our findings highlight that current LRMs largely encode interpretable processes, and interpretability itself can be a signal of prediction correctness.

Diffusion Models (DM) have democratized AI image generation through an iterative denoising process. Quantization is a major technique to alleviate the inference cost and reduce the size of DM denoiser networks. However, as denoisers evolve from variants of convolutional U-Nets toward newer Transformer architectures, it is of growing importance to understand the quantization sensitivity of different weight layers, operations and architecture types to performance. In this work, we address this challenge with Qua^2SeDiMo, a mixed-precision Post-Training Quantization framework that generates explainable insights on the cost-effectiveness of various model weight quantization methods for different denoiser operation types and block structures. We leverage these insights to make high-quality mixed-precision quantization decisions for a myriad of diffusion models ranging from foundational U-Nets to state-of-the-art Transformers. As a result, Qua^2SeDiMo can construct 3.4-bit, 3.9-bit, 3.65-bit and 3.7-bit weight quantization on PixArt-{alpha}, PixArt-{Sigma}, Hunyuan-DiT and SDXL, respectively. We further pair our weight-quantization configurations with 6-bit activation quantization and outperform existing approaches in terms of quantitative metrics and generative image quality.

Video generation models hold substantial potential in areas such as filmmaking. However, current video diffusion models need high computational costs and produce suboptimal results due to high complexity of video generation task. In this paper, we propose ConFiner, an efficient high-quality video generation framework that decouples video generation into easier subtasks: structure control and spatial-temporal refinement. It can generate high-quality videos with chain of off-the-shelf diffusion model experts, each expert responsible for a decoupled subtask. During the refinement, we introduce coordinated denoising, which can merge multiple diffusion experts' capabilities into a single sampling. Furthermore, we design ConFiner-Long framework, which can generate long coherent video with three constraint strategies on ConFiner. Experimental results indicate that with only 10\% of the inference cost, our ConFiner surpasses representative models like Lavie and Modelscope across all objective and subjective metrics. And ConFiner-Long can generate high-quality and coherent videos with up to 600 frames.

Recent advancements in the Generative Reward Model (GRM) have demonstrated its potential to enhance the reasoning abilities of LLMs through Chain-of-Thought (CoT) prompting. Despite these gains, existing implementations of GRM suffer from two critical limitations. First, CoT prompting is applied indiscriminately to all inputs regardless of their inherent complexity. This introduces unnecessary computational costs for tasks amenable to fast, direct inference. Second, existing approaches primarily rely on voting-based mechanisms to evaluate CoT outputs, which often lack granularity and precision in assessing reasoning quality. In this paper, we propose E-GRM, an efficient generative reward modeling framework grounded in model-internal uncertainty. E-GRM leverages the convergence behavior of parallel model generations to estimate uncertainty and selectively trigger CoT reasoning only when needed, without relying on handcrafted features or task-dependent signals. To improve reward fidelity, we introduce a lightweight discriminative scorer trained with a hybrid regression--ranking objective to provide fine-grained evaluation of reasoning paths. Experiments on multiple reasoning benchmarks show that E-GRM substantially reduces inference cost while consistently improving answer accuracy, demonstrating that model-internal uncertainty is an effective and general signal for efficient reasoning-aware reward modeling.

In recent years, the advent of foundation models (FM) for digital pathology has relied heavily on scaling the pre-training datasets and the model size, yielding large and powerful models. While it resulted in improving the performance on diverse downstream tasks, it also introduced increased computational cost and inference time. In this work, we explore the distillation of a large foundation model into a smaller one, reducing the number of parameters by several orders of magnitude. Leveraging distillation techniques, our distilled model, H0-mini, achieves nearly comparable performance to large FMs at a significantly reduced inference cost. It is evaluated on several public benchmarks, achieving 3rd place on the HEST benchmark and 5th place on the EVA benchmark. Additionally, a robustness analysis conducted on the PLISM dataset demonstrates that our distilled model reaches excellent robustness to variations in staining and scanning conditions, significantly outperforming other state-of-the art models. This opens new perspectives to design lightweight and robust models for digital pathology, without compromising on performance.

Computer Use Agents (CUAs) translate natural-language instructions into Graphical User Interface (GUI) actions such as clicks, keystrokes, and scrolls by relying on a Vision-Language Model (VLM) to interpret screenshots and predict grounded tool calls. However, grounding accuracy varies dramatically across VLMs, while current CUA systems typically route every action to a single fixed model regardless of difficulty. We propose Adaptive VLM Routing (AVR), a framework that inserts a lightweight semantic routing layer between the CUA orchestrator and a pool of VLMs. For each tool call, AVR estimates action difficulty from multimodal embeddings, probes a small VLM to measure confidence, and routes the action to the cheapest model whose predicted accuracy satisfies a target reliability threshold. For warm agents with memory of prior UI interactions, retrieved context further narrows the capability gap between small and large models, allowing many actions to be handled without escalation. We formalize routing as a cost--accuracy trade-off, derive a threshold-based policy for model selection, and evaluate AVR using ScreenSpot-Pro grounding data together with the OpenClaw agent routing benchmark. Across these settings, AVR projects inference cost reductions of up to 78\% while staying within 2 percentage points of an all-large-model baseline. When combined with the Visual Confused Deputy guardrail, AVR also escalates high-risk actions directly to the strongest available model, unifying efficiency and safety within a single routing framework. Materials are also provided Model, benchmark, and code: https://github.com/vllm-project/semantic-router.

Self-ensembling techniques with diverse reasoning paths such as Self-Consistency have demonstrated remarkable performance gains in text generation with Large Language Models (LLMs). However, such techniques depend on the availability of an accurate answer extraction process to aggregate across multiple outputs. Moreover, they acquire higher inference cost, in comparison to Greedy Decoding, due to generation of relatively higher number of output tokens. Research has shown that the free form text outputs from Self-Consistency can be aggregated reliably using LLMs to produce the final output. Additionally, recent advancements in LLM inference have demonstrated that usage of diverse exemplars in prompts have the ability to induce diversity in the LLM outputs. Such proven techniques can be easily extended to self-ensembling based approaches to achieve enhanced results in text generation. In this paper, we introduce PEDAL (Prompts based on Exemplar Diversity Aggregated using LLMs), a hybrid self-ensembling approach, that combines the strengths of diverse exemplar based prompts and LLM based aggregation to achieve improvement in overall performance. On the publicly available SVAMP and ARC datasets, our experiments reveal that PEDAL can achieve better accuracy than Greedy Decoding based strategies with lower inference cost compared to Self Consistency based approaches.

Knowledge distillation (KD) is widely used for compressing a teacher model to reduce its inference cost and memory footprint, by training a smaller student model. However, current KD methods for auto-regressive sequence models suffer from distribution mismatch between output sequences seen during training and those generated by the student during inference. To address this issue, we introduce Generalized Knowledge Distillation (GKD). Instead of solely relying on a fixed set of output sequences, GKD trains the student on its self-generated output sequences by leveraging feedback from the teacher on such sequences. Unlike supervised KD approaches, GKD also offers the flexibility to employ alternative loss functions between the student and teacher, which can be useful when the student lacks the expressivity to mimic the teacher's distribution. Furthermore, GKD facilitates the seamless integration of distillation with RL fine-tuning (RLHF). We demonstrate the efficacy of GKD for distilling auto-regressive language models on summarization, translation, and arithmetic reasoning tasks, and task-agnostic distillation for instruction-tuning.

Autoregressive video diffusion models (ARVDs) have emerged as a promising architecture for streaming video generation, paving the way for real-time interactive video generation and world modeling. Despite their potential, the substantial inference cost of ARVDs remains a major obstacle to practical deployment, making model quantization a natural direction for improving efficiency. However, quantization for ARVDs remains largely unexplored. Our empirical analysis shows that directly applying existing quantization schemes developed for standard diffusion transformers to ARVDs leads to suboptimal performance, revealing quantization behaviors that differ from those observed in bidirectional diffusion models. In this paper, we identify two critical challenges in quantizing ARVDs: (C1) Highly unbalanced frame-wise quantization sensitivity. Error accumulation during autoregressive generation can induce severely skewed quantization sensitivity across frames, following an exponential-like decay pattern. (C2) Prominent and heterogeneous outlier patterns in weights. Weight distributions exhibit pronounced outlier channels, whose patterns vary substantially across layer types and block depths. To address these issues, we propose Q-ARVD, a novel framework for accurate ARVD quantization. (S1) To tackle the highly unbalanced frame-wise sensitivity, Q-ARVD incorporates a final-quality aware frame-weighting mechanism into the quantization objective. (S2) To prevent heterogeneous outliers from degrading performance, Q-ARVD introduces an outlier-aware adaptive dual-scale quantization, which automatically detects the presence and quantity of outlier channels for an arbitrary layer, and isolates them to protect normal channels. Extensive experiments demonstrate the superiority of Q-ARVD.

"Talk short. Drop grammar. Save token." This caveman style is widely promoted as a way to cut inference cost, but whether it actually saves anything depends on which channel (the user's prompt or the model's response) is being compressed. We present Cavewoman, a two-channel evaluation protocol that scores every generation on task accuracy, realized per-item cost, and reference-text agreement against the model's unconstrained reference. We evaluate eight models on five datasets at five reduction levels, with both channels measured on the same items. Output compression cuts realized cost on most API models (1.4-2.4x per model, up to 3x in the best case) and on all four open-weight models under public-tier pricing. Input compression has the opposite effect, a strict lose-lose: it raises net cost rather than lowering it (~1.15x on the five-benchmark mean, up to 1.8x on the worst dataset and 2.7x under stronger compression), because models compensate with longer responses even as accuracy collapses. Under the same setting, surface text diverges from the unconstrained reference: on the non-reasoning models, roughly half of all generations are correct yet their surface text no longer entails the model's own unconstrained baseline generation. The divergence survives length-controlled re-scoring, multiple-comparisons correction, and replication under complementary semantic measures. Code and data are available at https://github.com/danielle34/cavewoman.

Multimodal recommendation aims to recommend user-preferred candidates based on her/his historically interacted items and associated multimodal information. Previous studies commonly employ an embed-and-retrieve paradigm: learning user and item representations in the same embedding space, then retrieving similar candidate items for a user via embedding inner product. However, this paradigm suffers from inference cost, interaction modeling, and false-negative issues. Toward this end, we propose a new MMGRec model to introduce a generative paradigm into multimodal recommendation. Specifically, we first devise a hierarchical quantization method Graph RQ-VAE to assign Rec-ID for each item from its multimodal and CF information. Consisting of a tuple of semantically meaningful tokens, Rec-ID serves as the unique identifier of each item. Afterward, we train a Transformer-based recommender to generate the Rec-IDs of user-preferred items based on historical interaction sequences. The generative paradigm is qualified since this model systematically predicts the tuple of tokens identifying the recommended item in an autoregressive manner. Moreover, a relation-aware self-attention mechanism is devised for the Transformer to handle non-sequential interaction sequences, which explores the element pairwise relation to replace absolute positional encoding. Extensive experiments evaluate MMGRec's effectiveness compared with state-of-the-art methods.

We introduce the Falcon series: 7B, 40B, and 180B parameters causal decoder-only models trained on a diverse high-quality corpora predominantly assembled from web data. The largest model, Falcon-180B, has been trained on over 3.5 trillion tokens of text--the largest openly documented pretraining run. Falcon-180B significantly outperforms models such as PaLM or Chinchilla, and improves upon concurrently developed models such as LLaMA 2 or Inflection-1. It nears the performance of PaLM-2-Large at a reduced pretraining and inference cost, making it, to our knowledge, one of the three best language models in the world along with GPT-4 and PaLM-2-Large. We report detailed evaluations, as well as a deep dive into the methods and custom tooling employed to pretrain Falcon. Notably, we report on our custom distributed training codebase, allowing us to efficiently pretrain these models on up to 4,096 A100s on cloud AWS infrastructure with limited interconnect. We release a 600B tokens extract of our web dataset, as well as the Falcon-7/40/180B models under a permissive license to foster open-science and accelerate the development of an open ecosystem of large language models.

To build an artificial neural network like the biological intelligence system, recent works have unified numerous tasks into a generalist model, which can process various tasks with shared parameters and do not have any task-specific modules. While generalist models achieve promising results on various benchmarks, they have performance degradation on some tasks compared with task-specialized models. In this work, we find that interference among different tasks and modalities is the main factor to this phenomenon. To mitigate such interference, we introduce the Conditional Mixture-of-Experts (Conditional MoEs) to generalist models. Routing strategies under different levels of conditions are proposed to take both the training/inference cost and generalization ability into account. By incorporating the proposed Conditional MoEs, the recently proposed generalist model Uni-Perceiver can effectively mitigate the interference across tasks and modalities, and achieves state-of-the-art results on a series of downstream tasks via prompt tuning on 1% of downstream data. Moreover, the introduction of Conditional MoEs still holds the generalization ability of generalist models to conduct zero-shot inference on new tasks, e.g., video-text retrieval and video caption. Code and pre-trained generalist models shall be released.

Recent work has shown that directly fine-tuning large language models (LLMs) for dense retrieval yields strong performance, but their substantial parameter counts make them computationally inefficient. While prior studies have revealed significant layer redundancy in LLMs for generative tasks, it remains unclear whether similar redundancy exists when these models are adapted for retrieval tasks, which require encoding entire sequences into fixed representations rather than generating tokens iteratively. To this end, we conduct a comprehensive analysis of layer redundancy in LLM-based dense retrievers. We find that, in contrast to generative settings, MLP layers are substantially more prunable, while attention layers remain critical for semantic aggregation. Building on this insight, we propose EffiR, a framework for developing efficient retrievers that performs large-scale MLP compression through a coarse-to-fine strategy (coarse-grained depth reduction followed by fine-grained width reduction), combined with retrieval-specific fine-tuning. Across diverse BEIR datasets and LLM backbones, EffiR achieves substantial reductions in model size and inference cost while preserving the performance of full-size models.

Mixture of expert (MoE) models are a promising approach to increasing model capacity without increasing inference cost, and are core components of many state-of-the-art language models. However, current MoE models typically use only few experts due to prohibitive training and inference cost. We propose Test-Time Model Merging (TTMM) which scales the MoE paradigm to an order of magnitude more experts and uses model merging to avoid almost any test-time overhead. We show that TTMM is an approximation of test-time training (TTT), which fine-tunes an expert model for each prediction task, i.e., prompt. TTT has recently been shown to significantly improve language models, but is computationally expensive. We find that performance of TTMM improves with more experts and approaches the performance of TTT. Moreover, we find that with a 1B parameter base model, TTMM is more than 100x faster than TTT at test-time by amortizing the cost of TTT at train-time. Thus, TTMM offers a promising cost-effective approach to scale test-time training.

Knowledge distillation (KD) is a common approach to compress a teacher model to reduce its inference cost and memory footprint, by training a smaller student model. However, in the context of autoregressive language models (LMs), we empirically find that larger teacher LMs might dramatically result in a poorer student. In response to this problem, we conduct a series of analyses and reveal that different tokens have different teaching modes, neglecting which will lead to performance degradation. Motivated by this, we propose a simple yet effective adaptive teaching approach (ATKD) to improve the KD. The core of ATKD is to reduce rote learning and make teaching more diverse and flexible. Extensive experiments on 8 LM tasks show that, with the help of ATKD, various baseline KD methods can achieve consistent and significant performance gains (up to +3.04% average score) across all model types and sizes. More encouragingly, ATKD can improve the student model generalization effectively.

Tangent Model Composition (TMC) is a method to combine component models independently fine-tuned around a pre-trained point. Component models are tangent vectors to the pre-trained model that can be added, scaled, or subtracted to support incremental learning, ensembling, or unlearning. Component models are composed at inference time via scalar combination, reducing the cost of ensembling to that of a single model. TMC improves accuracy by 4.2% compared to ensembling non-linearly fine-tuned models at a 2.5x to 10x reduction of inference cost, growing linearly with the number of component models. Each component model can be forgotten at zero cost, with no residual effect on the resulting inference. When used for continual fine-tuning, TMC is not constrained by sequential bias and can be executed in parallel on federated data. TMC outperforms recently published continual fine-tuning methods almost uniformly on each setting -- task-incremental, class-incremental, and data-incremental -- on a total of 13 experiments across 3 benchmark datasets, despite not using any replay buffer. TMC is designed for composing models that are local to a pre-trained embedding, but could be extended to more general settings.

Recent advancements in reasoning language models have demonstrated remarkable performance in complex tasks, but their extended chain-of-thought reasoning process increases inference overhead. While quantization has been widely adopted to reduce the inference cost of large language models, its impact on reasoning models remains understudied. In this study, we conduct the first systematic study on quantized reasoning models, evaluating the open-sourced DeepSeek-R1-Distilled Qwen and LLaMA families ranging from 1.5B to 70B parameters, and QwQ-32B. Our investigation covers weight, KV cache, and activation quantization using state-of-the-art algorithms at varying bit-widths, with extensive evaluation across mathematical (AIME, MATH-500), scientific (GPQA), and programming (LiveCodeBench) reasoning benchmarks. Our findings reveal that while lossless quantization can be achieved with W8A8 or W4A16 quantization, lower bit-widths introduce significant accuracy risks. We further identify model size, model origin, and task difficulty as critical determinants of performance. Contrary to expectations, quantized models do not exhibit increased output lengths. In addition, strategically scaling the model sizes or reasoning steps can effectively enhance the performance. All quantized models and codes will be open-sourced in https://github.com/ruikangliu/Quantized-Reasoning-Models.

Flow-based generative models have become a strong framework for high-quality generative modeling, yet pretrained models are rarely used in their vanilla conditional form: conditional samples without guidance often appear diffuse and lack fine-grained detail due to the smoothing effects of neural networks. Existing guidance techniques such as classifier-free guidance (CFG) improve fidelity but double the inference cost and typically reduce sample diversity. We introduce Momentum Guidance (MG), a new dimension of guidance that leverages the ODE trajectory itself. MG extrapolates the current velocity using an exponential moving average of past velocities and preserves the standard one-evaluation-per-step cost. It matches the effect of standard guidance without extra computation and can further improve quality when combined with CFG. Experiments demonstrate MG's effectiveness across benchmarks. Specifically, on ImageNet-256, MG achieves average improvements in FID of 36.68% without CFG and 25.52% with CFG across various sampling settings, attaining an FID of 1.597 at 64 sampling steps. Evaluations on large flow-based models like Stable Diffusion 3 and FLUX.1-dev further confirm consistent quality enhancements across standard metrics.

Token serves as the fundamental unit of computation in modern autoregressive models, and generation length directly influences both inference cost and reasoning performance. Despite its importance, existing approaches lack fine-grained length modeling, operating primarily at the coarse-grained sequence level. We introduce the Length Value Model (LenVM), a token-level framework that models the remaining generation length. By formulating length modeling as a value estimation problem and assigning a constant negative reward to each generated token, LenVM predicts a bounded, discounted return that serves as a monotone proxy for the remaining generation horizon. This formulation yields supervision that is annotation-free, dense, unbiased, and scalable. Experiments on LLMs and VLMs demonstrate LenVM provides a highly effective signal at inference time. On the LIFEBench exact length matching task, applying LenVM to a 7B model improves the length score from 30.9 to 64.8, significantly outperforming frontier closed-source models. Furthermore, LenVM enables continuous control over the trade off between performance and efficiency. On GSM8K at a budget of 200 tokens, LenVM maintains 63% accuracy compared to 6 percent for token budget baseline. It also accurately predicts total generation length from the prompt boundary. Finally, LenVM's token-level values offer an interpretable view of generation dynamics, revealing how specific tokens shift reasoning toward shorter or longer regimes. Results demonstrate that LenVM supports a broad range of applications and token length can be effectively modeled as a token-level value signal, highlighting the potential of LenVM as a general framework for length modeling and as a length-specific value signal that could support future RL training. Code is available at https://github.com/eric-ai-lab/Length-Value-Model.

Sparse Mixture-of-Experts (MoE) has been a successful approach for scaling multilingual translation models to billions of parameters without a proportional increase in training computation. However, MoE models are prohibitively large and practitioners often resort to methods such as distillation for serving. In this work, we investigate routing strategies at different granularity (token, sentence, task) in MoE models to bypass distillation. Experiments on WMT and a web-scale dataset suggest that task-level routing (task-MoE) enables us to extract smaller, ready-to-deploy sub-networks from large sparse models. On WMT, our task-MoE with 32 experts (533M parameters) outperforms the best performing token-level MoE model (token-MoE) by +1.0 BLEU on average across 30 language pairs. The peak inference throughput is also improved by a factor of 1.9x when we route by tasks instead of tokens. While distilling a token-MoE to a smaller dense model preserves only 32% of the BLEU gains, our sub-network task-MoE, by design, preserves all the gains with the same inference cost as the distilled student model. Finally, when scaling up to 200 language pairs, our 128-expert task-MoE (13B parameters) performs competitively with a token-level counterpart, while improving the peak inference throughput by a factor of 2.6x.

Processing video in vision-language models is expensive: each frame occupies hundreds of tokens, and inference cost scales with every frame and every repeated query. We introduce Video2LoRA, a method for parametric video internalization. A perceiver hypernetwork reads the intermediate representations produced layer-by-layer as a frozen VLM encodes a video, and generates a Low-Rank Adaptation (LoRA) adapter in a single forward pass. Unlike standard LoRA fine-tuning, which requires iterative gradient updates, Video2LoRA predicts these weights directly from the video. Trained for SmolVLM2 500M and 2.2B on video summarization and captioning, Video2LoRA enables the same frozen VLM to answer queries from the adapter alone, with zero visual tokens in its context at query time. Video2LoRA is statistically non-inferior and equivalent to direct video-in-context inference across all five captioning benchmarks at both model scales, and across seven of eight video question answering benchmark-scale pairings. Although trained only on 12 frames at 384px, it remains stable up to 1,024 frames and 1024px, where direct video-in-context inference often degenerates. Across this sweep, it reduces answer-time visual-token load by up to 1,500x and query TTFT by 6-80x, while preserving video-faithful outputs. We also find that independently generated adapters for non-overlapping video segments can compose in rank space, suggesting a path toward chunked long-video internalization.

Recent research shows the potential of enhancing the problem-solving ability of large language models (LLMs) through the use of external tools. However, prior work along this line depends on the availability of existing tools. In this work, we take an initial step towards removing this dependency by proposing a closed-loop framework, referred to as LLMs As Tool Makers (LATM), where LLMs create their own reusable tools for problem-solving. Our approach consists of two key phases: 1) tool making: an LLM acts as the tool maker that crafts tools for given tasks, where a tool is implemented as a Python utility function. 2) tool using: an LLM acts as the tool user, which applies the tool built by the tool maker for problem-solving. The tool user can be either the same or a different LLM from the tool maker. Tool-making enables an LLM to continually generate tools that can be applied to different requests so that future requests can call the corresponding APIs when beneficial for solving the tasks. Furthermore, the division of labor among LLMs for tool-making and tool-using phases introduces the opportunity to achieve cost effectiveness without degrading the quality of generated tools and problem solutions. For example, recognizing that tool-making demands more sophisticated capabilities than tool-using, we can apply a powerful yet resource-intensive model as the tool maker, and a lightweight while cost-effective model as the tool user. We validate the effectiveness of our approach across a variety of complex reasoning tasks, including Big-Bench tasks. With GPT-4 as the tool maker and GPT-3.5 as the tool user, LATM can achieve performance that is on par with using GPT-4 for both tool making and tool using, while the inference cost is significantly reduced.

Vision-language models improve perception by feeding increasingly long visual token sequences into language backbones, but the resulting inference cost raises a basic scaling question: as multimodal models grow, how many visual tokens are actually needed, and how should they be allocated under a fixed visual token budget? Existing training-free pruning methods typically answer this with one-shot proxies such as decoder attention, visual similarity, or conditional diversity. We argue that visual token pruning is better viewed as task-conditioned evidence search, especially under aggressive compression and across model scales. We propose F^3A, a training-free router for visual token pruning that operates before the language model consumes image tokens. F^3A builds lightweight question-conditioned cues, matches them to visual-grid tokens through frozen sparse sensing heads, and allocates a fixed vision token budget via coarse evidence localization, local refinement, coverage-preserving competition, and recovery of under-covered regions. It requires no model training, no extra LLM forward pass and preserves the original multimodal prompting and decoding pipeline.

Small, highly trained, open-source large language models are widely used due to their inference efficiency, but further improving their quality remains a challenge. Sparse upcycling is a promising approach that transforms a pretrained dense model into a Mixture-of-Experts (MoE) architecture, increasing the model's parameter count and quality. In this work, we compare the effectiveness of sparse upcycling against continued pretraining (CPT) across different model sizes, compute budgets, and pretraining durations. Our experiments show that sparse upcycling can achieve better quality, with improvements of over 20% relative to CPT in certain scenarios. However, this comes with a significant inference cost, leading to 40% slowdowns in high-demand inference settings for larger models. Our findings highlight the trade-off between model quality and inference efficiency, offering insights for practitioners seeking to balance model quality and deployment constraints.

The emergence of Mixture of Experts (MoE) LLMs has significantly advanced the development of language models. Compared to traditional LLMs, MoE LLMs outperform traditional LLMs by achieving higher performance with considerably fewer activated parameters. Despite this efficiency, their enormous parameter size still leads to high deployment costs. In this paper, we introduce a two-stage compression method tailored for MoE to reduce the model size and decrease the computational cost. First, in the inter-expert pruning stage, we analyze the importance of each layer and propose the Layer-wise Genetic Search and Block-wise KT-Reception Field with the non-uniform pruning ratio to prune the individual expert. Second, in the intra-expert decomposition stage, we apply the low-rank decomposition to further compress the parameters within the remaining experts. Extensive experiments on Qwen1.5-MoE-A2.7B, DeepSeek-V2-Lite, and Mixtral-8times7B demonstrate that our proposed methods can both reduce the model size and enhance inference efficiency while maintaining performance in various zero-shot tasks. The code will be available at https://github.com/xiaochengsky/MoEI-2.git

Despite the increasing use of large language models (LLMs) for context-grounded tasks like summarization and question-answering, understanding what makes an LLM produce a certain response is challenging. We propose Multi-Level Explanations for Generative Language Models (MExGen), a technique to provide explanations for context-grounded text generation. MExGen assigns scores to parts of the context to quantify their influence on the model's output. It extends attribution methods like LIME and SHAP to LLMs used in context-grounded tasks where (1) inference cost is high, (2) input text is long, and (3) the output is text. We conduct a systematic evaluation, both automated and human, of perturbation-based attribution methods for summarization and question answering. The results show that our framework can provide more faithful explanations of generated output than available alternatives, including LLM self-explanations. We open-source code for MExGen as part of the ICX360 toolkit: https://github.com/IBM/ICX360.

We introduce Arrow, a foundation model for zero-shot causal discovery on observational tabular data. Arrow factorizes a directed acyclic graph into an undirected skeleton and a topological order, guaranteeing acyclicity by construction. Given a new dataset, it uses a transformer-based architecture to contextualize variables within and across observations, then predicts skeleton edge probabilities and node order scores that together define a graph. Arrow is trained in a supervised fashion on synthetic datasets with ground-truth graphs, using an end-to-end differentiable directed edge composite likelihood induced by the skeleton-order factorization. The training distribution spans diverse graph families, functional forms, noise models, and dataset shapes. Across in- and out-of-distribution synthetic, semi-synthetic, and real datasets, Arrow matches or outperforms existing causal discovery methods at substantially lower inference cost than competitive alternatives. Our results demonstrate that large-scale pretraining on diverse synthetic data can yield zero-shot causal discovery models that are fast, accurate, and reusable on new datasets.

Evolutionary agentic systems intensify the trade-off between computational efficiency and reasoning capability by repeatedly invoking large language models (LLMs) during inference. This setting raises a central question: how can an agent dynamically select an LLM that is sufficiently capable for the current generation step while remaining computationally efficient? While model cascades offer a practical mechanism for balancing this trade-off, existing routing strategies typically rely on static heuristics or external controllers and do not explicitly account for model uncertainty. We introduce AdaptEvolve: Adaptive LLM Selection for Multi-LLM Evolutionary Refinement within an evolutionary sequential refinement framework that leverages intrinsic generation confidence to estimate real-time solvability. Empirical results show that confidence-driven selection yields a favourable Pareto frontier, reducing total inference cost by an average of 37.9% across benchmarks while retaining 97.5% of the upper-bound accuracy of static large-model baselines. Our code is available at https://github.com/raypretam/adaptive_llm_selection.

Large Language Model (LLM) agents have emerged as powerful tools for automating complex tasks by leveraging the reasoning and decision-making abilities of LLMs. However, a major bottleneck in current agent frameworks lies in the high inference cost of tool selection, especially in approaches like ReAct that repeatedly invoke the LLM to determine which tool to use at each step. In this work, we propose AutoTool, a novel graph-based framework that bypasses repeated LLM inference by exploiting a key empirical observation: tool usage inertia - the tendency of tool invocations to follow predictable sequential patterns. AutoTool constructs a directed graph from historical agent trajectories, where nodes represent tools and edges capture transition probabilities, effectively modeling the inertia in tool selection. It further integrates parameter-level information to refine tool input generation. By traversing this structured representation, AutoTool efficiently selects tools and their parameters with minimal reliance on LLM inference. Extensive experiments across diverse agent tasks demonstrate that AutoTool reduces inference costs by up to 30% while maintaining competitive task completion rates, offering a practical and scalable enhancement for inference-heavy frameworks. Our work highlights the promise of integrating statistical structure into LLM agent design for greater efficiency without sacrificing performance.

Modern recurrent layers are emerging as a promising path toward edge deployment of foundation models, especially in the context of large language models (LLMs). Compressing the whole input sequence in a finite-dimensional representation enables recurrent layers to model long-range dependencies while maintaining a constant inference cost for each token and a fixed memory requirement. However, the practical deployment of LLMs in resource-limited environments often requires further model compression, such as quantization and pruning. While these techniques are well-established for attention-based models, their effects on recurrent layers remain underexplored. In this preliminary work, we focus on post-training quantization for recurrent LLMs and show that Mamba models exhibit the same pattern of outlier channels observed in attention-based LLMs. We show that the reason for the difficulty of quantizing SSMs is caused by activation outliers, similar to those observed in transformer-based LLMs. We report baseline results for post-training quantization of Mamba that do not take into account the activation outliers and suggest first steps for outlier-aware quantization.

Scaling laws have been recently employed to derive compute-optimal model size (number of parameters) for a given compute duration. We advance and refine such methods to infer compute-optimal model shapes, such as width and depth, and successfully implement this in vision transformers. Our shape-optimized vision transformer, SoViT, achieves results competitive with models that exceed twice its size, despite being pre-trained with an equivalent amount of compute. For example, SoViT-400m/14 achieves 90.3% fine-tuning accuracy on ILSRCV2012, surpassing the much larger ViT-g/14 and approaching ViT-G/14 under identical settings, with also less than half the inference cost. We conduct a thorough evaluation across multiple tasks, such as image classification, captioning, VQA and zero-shot transfer, demonstrating the effectiveness of our model across a broad range of domains and identifying limitations. Overall, our findings challenge the prevailing approach of blindly scaling up vision models and pave a path for a more informed scaling.

In large vision-language models (LVLMs), images serve as inputs that carry a wealth of information. As the idiom "A picture is worth a thousand words" implies, representing a single image in current LVLMs can require hundreds or even thousands of tokens. This results in significant computational costs, which grow quadratically as input image resolution increases, thereby severely impacting the efficiency of both training and inference. Previous approaches have attempted to reduce the number of image tokens either before or within the early layers of LVLMs. However, these strategies inevitably result in the loss of crucial image information, ultimately diminishing model performance. To address this challenge, we conduct an empirical study revealing that all visual tokens are necessary for LVLMs in the shallow layers, and token redundancy progressively increases in the deeper layers of the model. To this end, we propose PyramidDrop, a visual redundancy reduction strategy for LVLMs to boost their efficiency in both training and inference with neglectable performance loss. Specifically, we partition the LVLM into several stages and drop part of the image tokens at the end of each stage with a pre-defined ratio, creating pyramid-like visual tokens across model layers. The dropping is based on a lightweight similarity calculation with a negligible time overhead. Extensive experiments demonstrate that PyramidDrop can achieve a 40% training time and 55% inference FLOPs acceleration of LLaVA-NeXT with comparable performance. Besides, the PyramidDrop could also serve as a plug-and-play strategy for inference acceleration without training, with better performance and lower inference cost than counterparts. We hope that the insights and approach introduced by PyramidDrop will inspire future research to further investigate the role of image tokens in LVLMs.

Omnimodal large language models (Omni-LLMs) show strong capability in audio-video understanding, but their practical deployment remains limited by high inference cost of long video streams and dense audio sequences. Despite recent progress, existing compression methods for Omni-LLMs typically rely on fixed or native compression units, which can disrupt cross-modal correspondence and the complementary information required for audio-video reasoning, making it difficult to improve inference efficiency while stably preserving performance. To address this, we propose OmniRefine, a training-free two-stage framework for efficient audio-visual token compression in Omni-LLMs. First, Correspondence-Preserving Chunk Refinement refines native chunk boundaries into cross-modally aligned compression units through frame-audio similarity and dynamic programming. Second, Modality-Aware Cooperative Compression jointly compresses video and audio tokens within each refined unit to reduce redundancy while preserving critical evidence. Extensive experiments show that OmniRefine achieves a better efficiency-performance trade-off than strong baselines and maintains stable performance under lower compression ratios. On WorldSense, it still reaches 46.7% accuracy at a 44% token retention ratio, nearly matching the full-token baseline. The code and interface will be released to facilitate further research.

Speculative decoding accelerates large language model inference without changing the target distribution, but no single drafting structure performs best across real-world workloads. Autoregressive multi-token prediction (MTP) is a lightweight, stable proposal mechanism, whereas block-parallel diffusion amortizes drafting latency over much longer candidate sequences; the better choice depends strongly on the output distribution. We present AngelSpec, a unified training framework for MTP and block-parallel speculative decoding that addresses this heterogeneity at three levels. At the training level, rather than fitting one universal drafter to a uniform data mixture, we co-specialize structure and data: the MTP drafter is trained on diverse conversational data for high-entropy open-ended chat, and the block-diffusion drafter on code and mathematics data for longer predictable continuations. At the architecture level, we propose DFly, a block-diffusion framework combining a hybrid target-conditioning backbone with a predecessor-conditioned autoregressive head, improving target-feature utilization and intra-block dependency modeling while keeping generation parallel. At the inference level, both acceptance length and verification cost vary with domain, request, online load, and hardware, so DFly treats verification as a shared batch-level resource: it reallocates compute toward high-confidence prefixes across requests and combines expected utility with a profiled cost model to adapt verification depth online. Across the Hy3 series, DFly raises the average accepted length on Hy3-A21B by roughly 30% and attains the highest average throughput at every tested concurrency from 4 to 64, a 1.98-2.40x speedup over autoregressive decoding and 10.5-11.8% higher throughput than DFlash. We release AngelSpec to support training and extending these methods.

We introduce Humans-Junior, a 3.8B model that matches GPT-4o on the FACTS Grounding public subset within a pm 5 pp equivalence margin. Results. On Q1--Q500 under identical judges, GPT-4o scores 73.5% (95% CI 69.5--77.2) and Humans-Junior 72.7% (95% CI 68.7--76.5); the paired difference is 0.8 pp (bootstrap 95% CI -3.1 to +4.7; permutation p = 0.72; Cohen's d = 0.023). TOST establishes equivalence at pm 5 pp (not at pm 3 pp). When purchased as managed APIs, Humans-Junior's base model (Phi-3.5-mini-instruct) is approx 19times less expensive than GPT-4o on Microsoft AI Foundry pricing; self-hosted or edge deployments can drive incremental inference cost toward zero. Measured vs estimated pricing sources are tabulated in Appendix E. Method. Our approach combines minimal directed "Exoskeleton Reasoning" scaffolds with behavioral fine-tuning that teaches protocol compliance (epistemic discipline) rather than domain answers. Fine-tuning alone adds little; combined, they synergize (+17.7 pp, p < 0.001) and reduce variance (approx 25%). In prompt-only settings on frontier models (Q1--Q100; non-comparable), directed reasoning improved GPT-4o by +11.8 pp to 85.3% and Gemini-2.5-Pro by +5.0 pp to 93.3% (baseline 88.3%, n = 100); see Section~5. TL;DR. A 3.8B model achieves GPT-4o-level FACTS accuracy (equivalent within pm 5 pp on Q1--Q500). Cloud pricing shows approx 19times lower cost versus GPT-4o, and self-hosted/edge deployments can approach zero marginal cost. Pricing sources are listed in Appendix E. Frontier prompt-only gains (Q1--Q100; non-comparable) and optimized-prompt exploratory results under earlier judges are summarized in Appendix F. Keywords: Small Language Models, Factual Grounding, Directed Reasoning, Fine-Tuning, Model Alignment, Cost-Efficient AI

Chain-of-Thought (CoT) reasoning has demonstrated remarkable effectiveness in enhancing the reasoning abilities of large language models (LLMs). However, its efficiency remains a challenge due to the generation of excessive intermediate reasoning tokens, which introduce semantic redundancy and overly detailed reasoning steps. Moreover, computational expense and latency are significant concerns, as the cost scales with the number of output tokens, including those intermediate steps. In this work, we observe that most CoT tokens are unnecessary, and retaining only a small portion of them is sufficient for producing high-quality responses. Inspired by this, we propose HAWKEYE, a novel post-training and inference framework where a large model produces concise CoT instructions to guide a smaller model in response generation. HAWKEYE quantifies redundancy in CoT reasoning and distills high-density information via reinforcement learning. By leveraging these concise CoTs, HAWKEYE is able to expand responses while reducing token usage and computational cost significantly. Our evaluation shows that HAWKEYE can achieve comparable response quality using only 35% of the full CoTs, while improving clarity, coherence, and conciseness by approximately 10%. Furthermore, HAWKEYE can accelerate end-to-end reasoning by up to 3.4x on complex math tasks while reducing inference cost by up to 60%. HAWKEYE will be open-sourced and the models will be available soon.

Data curation has shifted the quality-compute frontier for language-model and contrastive image-text pretraining, but its role for vision-language models (VLMs) is far less established. We ask how far data curation alone can take VLM performance, holding architecture, training recipe, and compute fixed and varying only the training data. Our pipeline, applied to the MAmmoTH-VL single-image subset, lifts performance by +11.7pp on average across 20 public VLM benchmarks (spanning grounding, VQA, OCR/documents, captioning, spatial/3D, counting, charts, math, brand-ID, and multi-image reasoning) and by +11.3pp on average across all nine capability axes of DatBench, our high-fidelity VLM eval suite. At 2B, our curated model surpasses InternVL3.5-2B by 9.9pp at ~17x less training compute and closes the gap to Qwen3-VL-2B to within 1.8pp at ~87x less compute, from pretraining alone. Beyond accuracy, curation delivers four further properties: (1) Reliability: per-capability std across training seeds drops by ~67% and the lift survives a 4k-to-16k context-length sweep; (2) OOD generalization: the 9-eval OOD average rises by +7.2pp, and multi-image BLINK rises by +3.09pp despite single-image-only training, with Visual Correspondence gaining +11.8pp; (3) Behavioral gains beyond benchmarks: across ~1,100 open-ended queries the curated 2B is more honest and more specific than the matched-compute baseline, and more concise and less refusal-prone than a frontier 2B reference; (4) Pareto-dominance on inference cost: at every scale (1B, 2B, 4B) the curated model raises accuracy while lowering response FLOPs vs. the matched-compute baseline, and the curated 4B matches near-frontier accuracy at 3.3x lower response FLOPs than Qwen3-VL-4B. Data curation is a high-leverage tool for building better VLMs, reaching near-frontier accuracy at up to ~150x less training compute.

Graph-based Multi-Agent Systems (MAS) enable complex cyclic workflows but suffer from inefficient static model allocation, where deploying strong models uniformly wastes computation on trivial sub-tasks. We propose CASTER (Context-Aware Strategy for Task Efficient Routing), a lightweight router for dynamic model selection in graph-based MAS. CASTER employs a Dual-Signal Router that combines semantic embeddings with structural meta-features to estimate task difficulty. During training, the router self-optimizes through a Cold Start to Iterative Evolution paradigm, learning from its own routing failures via on-policy negative feedback. Experiments using LLM-as-a-Judge evaluation across Software Engineering, Data Analysis, Scientific Discovery, and Cybersecurity demonstrate that CASTER reduces inference cost by up to 72.4% compared to strong-model baselines while matching their success rates, and consistently outperforms both heuristic routing and FrugalGPT across all domains.

In search scenarios, user experience can be hindered by erroneous queries due to typos, voice errors, or knowledge gaps. Therefore, query correction is crucial for search engines. Current correction models, usually small models trained on specific data, often struggle with queries beyond their training scope or those requiring contextual understanding. While the advent of Large Language Models (LLMs) offers a potential solution, they are still limited by their pre-training data and inference cost, particularly for complex queries, making them not always effective for query correction. To tackle these, we propose Trigger^3, a large-small model collaboration framework that integrates the traditional correction model and LLM for query correction, capable of adaptively choosing the appropriate correction method based on the query and the correction results from the traditional correction model and LLM. Trigger^3 first employs a correction trigger to filter out correct queries. Incorrect queries are then corrected by the traditional correction model. If this fails, an LLM trigger is activated to call the LLM for correction. Finally, for queries that no model can correct, a fallback trigger decides to return the original query. Extensive experiments demonstrate Trigger^3 outperforms correction baselines while maintaining efficiency.

One of the guiding principles for designing AI-based weather forecasting systems is to embed physical constraints as inductive priors in the neural network architecture. A popular prior is locality, where the atmospheric data is processed with local neural interactions, like 3D convolutions or 3D local attention windows as in Pangu-Weather. On the other hand, some works have shown great success in weather forecasting without this locality principle, at the cost of a much higher parameter count. In this paper, we show that the 3D local processing in Pangu-Weather is computationally sub-optimal. We design ArchesWeather, a transformer model that combines 2D attention with a column-wise attention-based feature interaction module, and demonstrate that this design improves forecasting skill. ArchesWeather is trained at 1.5° resolution and 24h lead time, with a training budget of a few GPU-days and a lower inference cost than competing methods. An ensemble of four of our models shows better RMSE scores than the IFS HRES and is competitive with the 1.4° 50-members NeuralGCM ensemble for one to three days ahead forecasting. Our code and models are publicly available at https://github.com/gcouairon/ArchesWeather.

High-fidelity 3D object synthesis remains significantly more challenging than 2D image generation due to the unstructured nature of mesh data and the cubic complexity of dense volumetric grids. Existing two-stage pipelines-compressing meshes with a VAE (using either 2D or 3D supervision), followed by latent diffusion sampling-often suffer from severe detail loss caused by inefficient representations and modality mismatches introduced in VAE. We introduce Sparc3D, a unified framework that combines a sparse deformable marching cubes representation Sparcubes with a novel encoder Sparconv-VAE. Sparcubes converts raw meshes into high-resolution (1024^3) surfaces with arbitrary topology by scattering signed distance and deformation fields onto a sparse cube, allowing differentiable optimization. Sparconv-VAE is the first modality-consistent variational autoencoder built entirely upon sparse convolutional networks, enabling efficient and near-lossless 3D reconstruction suitable for high-resolution generative modeling through latent diffusion. Sparc3D achieves state-of-the-art reconstruction fidelity on challenging inputs, including open surfaces, disconnected components, and intricate geometry. It preserves fine-grained shape details, reduces training and inference cost, and integrates naturally with latent diffusion models for scalable, high-resolution 3D generation.

The reasoning-capable large language models (LLMs) demonstrate strong performance on complex reasoning tasks but often suffer from overthinking, generating unnecessarily long chain-of-thought (CoT) reasoning paths for easy reasoning questions, thereby increasing inference cost and latency. Recent approaches attempt to address this challenge by manually deciding when to apply long or short reasoning. However, they lack the flexibility to adapt CoT length dynamically based on question complexity. In this paper, we propose Auto Long-Short Reasoning (AutoL2S), a dynamic and model-agnostic framework that enables LLMs to dynamically compress their generated reasoning path based on the complexity of the reasoning question. AutoL2S enables a learned paradigm, in which LLMs themselves can decide when longer reasoning is necessary and when shorter reasoning suffices, by training on data annotated with our proposed method, which includes both long and short CoT paths and a special <EASY> token. We then use <EASY> token to indicate when the model can skip generating lengthy CoT reasoning. This proposed annotation strategy can enhance the LLMs' ability to generate shorter CoT reasoning paths with improved quality after training. Extensive evaluation results show that AutoL2S reduces the length of reasoning generation by up to 57% without compromising performance, demonstrating the effectiveness of AutoL2S for scalable and efficient LLM reasoning.

The training and inference of large language models (LLMs) are together a costly process that transports knowledge from raw data to meaningful computation. Inspired by the memory hierarchy of the human brain, we reduce this cost by equipping LLMs with explicit memory, a memory format cheaper than model parameters and text retrieval-augmented generation (RAG). Conceptually, with most of its knowledge externalized to explicit memories, the LLM can enjoy a smaller parameter size, training cost, and inference cost, all proportional to the amount of remaining "abstract knowledge". As a preliminary proof of concept, we train from scratch a 2.4B LLM, which achieves better performance than much larger LLMs as well as RAG models, and maintains higher decoding speed than RAG. The model is named Memory^3, since explicit memory is the third form of memory in LLMs after implicit memory (model parameters) and working memory (context key-values). We introduce a memory circuitry theory to support the externalization of knowledge, and present novel techniques including a memory sparsification mechanism that makes storage tractable and a two-stage pretraining scheme that facilitates memory formation.

Model routing chooses which language model to use for each query. By sending easy queries to cheaper models and hard queries to stronger ones, it can significantly reduce inference cost while maintaining high accuracy. However, most existing routers treat this as a fixed choice among a small set of models, which makes them hard to adapt to new models or changing budget constraints. In this paper, we propose SCOPE (Scalable and Controllable Outcome Performance Estimator), a routing framework that goes beyond model selection by predicting their cost and performance. Trained with reinforcement learning, SCOPE makes reasoning-based predictions by retrieving how models behave on similar problems, rather than relying on fixed model names, enabling it to work with new, unseen models. Moreover, by explicitly predicting how accurate and how expensive a model will be, it turns routing into a dynamic decision problem, allowing users to easily control the trade-off between accuracy and cost. Experiments show that SCOPE is more than just a cost-saving tool. It flexibly adapts to user needs: it can boost accuracy by up to 25.7% when performance is the priority, or cut costs by up to 95.1% when efficiency matters most.

Frontier LLM agents are automating many business tasks, but their high inference cost makes large-scale deployment unsustainable. Small language models (SLMs) offer a cheaper alternative, yet they typically fall short when swapped into a harness designed for a frontier LLM. We show that for many routine business tasks, SLM agents can match LLM performance at 90% lower cost, when paired with an adapted harness that can be automatically discovered by a meta agent. The key insight is that much of the task difficulty is shared across instances and can be lifted from the model into the harness via tailored instructions, tools, and orchestration loops. To study this systematically, we create a framework that maps agent failure modes to harness adaptation strategies, and build a harness optimizer that automatically discovers effective adaptations from failure trajectories. Across seven business-oriented agentic tasks and three SLM families, we found optimized harnesses significantly improve performance on 16 of 21 task-SLM pairs, with seven pairs closing the SLM-LLM performance gap and the best SLM agent recovering 89.7% of LLM performance at 4% of the cost. Our analysis further shows that adaptation works best for tasks with more repetitive workflows and for SLMs with sufficient base capabilities. Together, these results suggest that harness adaptation can expand the practical deployment range of SLM agents in routine business tasks.

A pervasive intuition holds that vision-language models (VLMs) are most trustworthy when their attention maps look sharp: concentrated attention on the queried region should imply a confident, calibrated answer. We test this Attention-Confidence Assumption directly. We instrument three open-weight VLM families (LLaVA-1.5, PaliGemma, Qwen2-VL; 3-7B parameters) with a unified mechanistic pipeline -- the VLM Reliability Probe (VRP) -- that compares attention structure, generation dynamics, and hidden-state geometry against a single correctness label. Three results emerge. (i) Attention structure is a near-zero predictor of correctness (R_pb(C_k,y)=0.001, 95% CI [-0.034,0.036]; R_pb(H_s,y)=-0.012, [-0.047,0.024] on a pooled n=3,090 split), even though attention remains causally necessary for feature extraction (top-30% patch masking drops accuracy by 8.2-11.3 pp, p<0.001). (ii) Reliability becomes legible later in the computation: a single hidden-state linear probe reaches AUROC>0.95 on POPE for two of three families, and self-consistency at K=10 is the strongest behavioral predictor we measure at 10x inference cost (R_pb=0.43). (iii) Causal neuron-level ablations expose a sharp architectural split with direct monitor-design implications: late-fusion LLaVA concentrates reliability in a fragile late bottleneck (-8.3 pp object-identification accuracy after top-5 probe-neuron ablation), whereas early-fusion PaliGemma and Qwen2-VL distribute it widely and absorb destruction of ~50% of their peak-layer hidden dimension with <=1 pp degradation. The takeaway is narrow but consequential: in 3-7B VLMs, reliability is read more reliably off hidden-state geometry, layer-wise margin formation, and sparse late-layer circuits than off attention-map sharpness.

Large language models (LLMs) are increasingly deployed in multi-turn dialogue settings, but their behavior is still bottlenecked by fixed context windows and naive memory strategies. Replaying the full conversation at every turn is simple but expensive, while static summarization or recency-only heuristics often erase safety-critical user details. We present Adaptive Focus Memory (AFM), a dynamic context manager that assigns each past message one of three fidelity levels -- FULL, COMPRESSED, or PLACEHOLDER -- based on semantic similarity to the current query, half-life recency weighting, and importance classification. AFM packs messages chronologically under a strict token budget, preferring high fidelity for the most relevant turns while aiming to preserve a cheap trace of the dialogue. In a safety-oriented benchmark involving a user with a severe peanut allergy planning a trip to Thailand, AFM retains the allergy across both short and medium-length conversations, matches the safety performance of naive replay, and cuts average token usage by 66% relative to a replay baseline. We release a modular Python implementation of AFM designed for OpenAI-compatible APIs and offline operation, enabling practitioners to reduce inference cost without sacrificing safety or factual continuity in the evaluated scenario.

Aligning diffusion models to downstream tasks often requires finetuning new models or gradient-based guidance at inference time to enable sampling from the reward-tilted posterior. In this work, we explore a simple inference-time gradient-free guidance approach, called controlled denoising (CoDe), that circumvents the need for differentiable guidance functions and model finetuning. CoDe is a blockwise sampling method applied during intermediate denoising steps, allowing for alignment with downstream rewards. Our experiments demonstrate that, despite its simplicity, CoDe offers a favorable trade-off between reward alignment, prompt instruction following, and inference cost, achieving a competitive performance against the state-of-the-art baselines. Our code is available at: https://github.com/anujinho/code.

Structured state-space models (SSMs) such as S4, stemming from the seminal work of Gu et al., are gaining popularity as effective approaches for modeling sequential data. Deep SSMs demonstrate outstanding performance across a diverse set of domains, at a reduced training and inference cost compared to attention-based transformers. Recent developments show that if the linear recurrence powering SSMs allows for multiplicative interactions between inputs and hidden states (e.g. GateLoop, Mamba, GLA), then the resulting architecture can surpass in both in accuracy and efficiency attention-powered foundation models trained on text, at scales of billion parameters. In this paper, we give theoretical grounding to this recent finding using tools from Rough Path Theory: we show that when random linear recurrences are equipped with simple input-controlled transitions (selectivity mechanism), then the hidden state is provably a low-dimensional projection of a powerful mathematical object called the signature of the input -- capturing non-linear interactions between tokens at distinct timescales. Our theory not only motivates the success of modern selective state-space models such as Mamba but also provides a solid framework to understand the expressive power of future SSM variants.

Knowledge distillation (KD) is widely used for compressing a teacher model to a smaller student model, reducing its inference cost and memory footprint while preserving model capabilities. However, current KD methods for auto-regressive sequence models (e.g., large language models) suffer from missing a standardized objective function. Moreover, the recent use of student-generated outputs to address training-inference mismatches has significantly escalated computational costs. To tackle these issues, we introduce DistiLLM, a more effective and efficient KD framework for auto-regressive language models. DistiLLM comprises two components: (1) a novel skew Kullback-Leibler divergence loss, where we unveil and leverage its theoretical properties, and (2) an adaptive off-policy approach designed to enhance the efficiency in utilizing student-generated outputs. Extensive experiments, including instruction-following tasks, demonstrate the effectiveness of DistiLLM in building high-performing student models while achieving up to 4.3times speedup compared to recent KD methods.

The demand for efficient deployment of large language models (LLMs) has driven interest in quantization, which reduces inference cost, and parameter-efficient fine-tuning (PEFT), which lowers training overhead. This motivated the development of quantization-aware PEFT to produce accurate yet efficient quantized models. In this setting, reducing quantization error prior to fine-tuning is crucial for achieving high model accuracy. However, existing methods that rely on low-rank adaptation suffer from limited representational capacity. Recent Fourier-related transform (FT)-based adapters offer greater representational power than low-rank adapters, but their direct integration into quantized models often results in ineffective error reduction and increased computational overhead. To overcome these limitations, we propose QWHA, a method that integrates FT-based adapters into quantized models by employing the Walsh-Hadamard Transform (WHT) as the transform kernel, together with a novel adapter initialization scheme incorporating adaptive parameter selection and value refinement. We demonstrate that QWHA effectively mitigates quantization errors while facilitating fine-tuning, and that its design substantially reduces computational cost. Experimental results show that QWHA consistently outperforms baselines in low-bit quantization accuracy and achieves significant training speedups over existing FT-based adapters. The code is available at https://github.com/vantaa89/qwha.

Zero-shot text-to-speech (TTS) now supports interactive assistants, personalized media, and accessibility tools. All TTS systems require faithful linguistic rendering, consistent speaker identity, and low-latency response. Yet compact streaming systems must preserve sufficient acoustic detail in a predictable low-rate latent sequence, while iterative diffusion sampling and classifier-free guidance multiply inference cost at every autoregressive step. To strike a balance between high-fidelity synthesis and low-latency inference, we present CuteTTS, a compact continuous-autoregressive TTS system. It combines semantically aligned causal VAE latents with patch-level autoregression, explicit speaker conditioning, and a bidirectional flow-matching head. We further introduce guidance-step distillation, which absorbs classifier-free guidance and multiple solver steps into a single interval-conditioned student. Evaluations on LibriSpeech and Seed-TTS-Eval demonstrate competitive intelligibility and speaker similarity in zero-shot voice cloning, while distillation lowers first-audio latency by 23.3% and real-time factor by 40.8% relative to the base model with comparable objective and subjective quality. These results provide a practical path toward continuous-autoregressive TTS that reconciles high-fidelity generation with the latency demands of real-time interaction.

Generative models trained using self-supervision of tokenized electronic health record (EHR) timelines show promise for clinical outcome prediction. This is typically done using Monte Carlo simulation for future patient trajectories. However, existing approaches suffer from three key limitations: sparse estimate distributions that poorly differentiate patient risk levels, extreme computational costs, and high sampling variance. We propose two new estimators: the Sum of Conditional Outcome Probability Estimator (SCOPE) and Risk Estimation from Anticipated Conditional Hazards (REACH), that leverage next-token probability distributions discarded by standard Monte Carlo. We prove both estimators are unbiased and that REACH guarantees variance reduction over Monte Carlo sampling for any model and outcome. Empirically, on hospital mortality prediction in MIMIC-IV using the ETHOS-ARES framework, SCOPE and REACH match 100-sample Monte Carlo performance using only 10-11 samples (95% CI: [9,11]), representing a ~10x reduction in inference cost without degrading calibration. For ICU admission prediction, efficiency gains are more modest (~1.2x), which we attribute to the outcome's lower "spontaneity," a property we characterize theoretically and empirically. These methods substantially improve the feasibility of deploying generative EHR models in resource-constrained clinical settings.

Recent advances in Automatic Speech Recognition (ASR) have demonstrated remarkable accuracy and robustness in diverse audio applications, such as live transcription and voice command processing. However, deploying these models on resource constrained edge devices (e.g., IoT device, wearables) still presents substantial challenges due to strict limits on memory, compute and power. Quantization, particularly Post-Training Quantization (PTQ), offers an effective way to reduce model size and inference cost without retraining. Despite its importance, the performance implications of various advanced quantization methods and bit-width configurations on ASR models remain unclear. In this work, we present a comprehensive benchmark of eight state-of-the-art (SOTA) PTQ methods applied to two leading edge-ASR model families, Whisper and Moonshine. We systematically evaluate model performances (i.e., accuracy, memory I/O and bit operations) across seven diverse datasets from the open ASR leaderboard, analyzing the impact of quantization and various configurations on both weights and activations. Built on an extension of the LLM compression toolkit, our framework integrates edge-ASR models, diverse advanced quantization algorithms, a unified calibration and evaluation data pipeline, and detailed analysis tools. Our results characterize the trade-offs between efficiency and accuracy, demonstrating that even 3-bit quantization can succeed on high capacity models when using advanced PTQ techniques. These findings provide valuable insights for optimizing ASR models on low-power, always-on edge devices.

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3times faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20% faster to infer. In addition, we introduce an FP8-based training recipe and show that it can achieve on par results with BF16-based training. This recipe is used to train the 56B model. All Nemotron-H models will be released, with support in Hugging Face, NeMo, and Megatron-LM.

Ternary weight quantization (e.g., BitNet b1.58) offers a promising path to mitigate the memory bandwidth bottleneck in Large Language Model (LLM) inference. However, conventional compute platforms lack native support for ternary-weight arithmetic, often relying on inefficient dequantization. Lookup table (LUT)-based hardware architectures provide an effective alternative by replacing multiplications with conditional additions, but their design space remains largely unexplored. Existing designs rely on heuristic parameter selection, lacking a systematic understanding of the architectural trade-offs. This work addresses this gap by formalizing the design space of ternary LUT-based accelerators and presenting an open-source hardware generator coupled with an analytical cost model, validated against synthesis in TSMC 16nm technology. By spanning the full architectural space, this framework not only enables rapid design space exploration but also establishes a common footing for fair cross-design evaluation, which was previously hindered by inconsistent instantiations across published accelerators. Using this framework, we challenge several assumptions and design choices in recent literature. We demonstrate that the optimal architecture is fundamentally governed by the activation data type: while LUT-based reuse offers significant gains for high-cost arithmetic (e.g., FP16), it yields diminishing returns for small integer types. Furthermore, we show that maximizing core size consistently improves area density compared to highly tiled approaches. Our optimized designs achieve a 2.2x area reduction compared to multiplier-based baselines. Moreover, by benchmarking state-of-the-art implementations against our model, we reveal that correcting suboptimal parameters yields up to a 1.2x area improvement.

Large Language Models (LLMs) deliver state-of-the-art performance across many tasks but impose high computational and memory costs, limiting their deployment in resource-constrained or real-time settings. To address this, we propose HierRouter, a hierarchical routing approach that dynamically assembles inference pipelines from a pool of specialized, lightweight language models. Formulated as a finite-horizon Markov Decision Process (MDP), our approach trains a Proximal Policy Optimization (PPO)-based reinforcement learning agent to iteratively select which models to invoke at each stage of multi-hop inference. The agent conditions on the evolving context and accumulated cost to make context-aware routing decisions. Experiments with three open-source candidate LLMs across six benchmarks, including QA, code generation, and mathematical reasoning, show that HierRouter improves response quality by up to 2.4x compared to using individual models independently, while incurring only a minimal additional inference cost on average. These results highlight the promise of hierarchical routing for cost-efficient, high-performance LLM inference. All codes can be found here https://github.com/ Nikunj-Gupta/hierouter.

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for open-world robot manipulation, but their practical deployment is often constrained by cost: billion-scale VLM backbones and iterative diffusion/flow-based action heads incur high latency and compute, making real-time control expensive on commodity hardware. We present A1, a fully open-source and transparent VLA framework designed for low-cost, high-throughput inference without sacrificing manipulation success; Our approach leverages pretrained VLMs that provide implicit affordance priors for action generation. We release the full training stack (training code, data/data-processing pipeline, intermediate checkpoints, and evaluation scripts) to enable end-to-end reproducibility. Beyond optimizing the VLM alone, A1 targets the full inference pipeline by introducing a budget-aware adaptive inference scheme that jointly accelerates the backbone and the action head. Specifically, we monitor action consistency across intermediate VLM layers to trigger early termination, and propose Inter-Layer Truncated Flow Matching that warm-starts denoising across layers, enabling accurate actions with substantially fewer effective denoising iterations. Across simulation benchmarks (LIBERO, VLABench) and real robots (Franka, AgiBot), A1 achieves state-of-the-art success rates while significantly reducing inference cost (e.g., up to 72% lower per-episode latency for flow-matching inference and up to 76.6% backbone computation reduction with minor performance degradation). On RoboChallenge, A1 achieves an average success rate of 29.00%, outperforming baselines including pi0(28.33%), X-VLA (21.33%), and RDT-1B (15.00%).

Predominant LLMs focus on high-resource languages while leaving low-resource languages, particularly those in Southeast Asia (SEA), underrepresented. In addition, those models are general-purpose and pay limited attention to the e-commerce domain. To overcome these limitations, we introduce Compass-v2, a lightweight Mixture-of-Experts (MoE) model specifically designed for Southeast Asian languages and e-commerce applications. To balance model performance and inference cost, the model is designed with 30B total parameters and 5B active parameters, incorporating both fine-grained and shared expert modules. To enhance multilingual performance, we curated and constructed a high-quality, industry-leading SEA dataset, to the best of our knowledge. To boost performance in the e-commerce domain, we built a dataset comprising hundreds of billions of tokens, sourced through external data mining and internal platform collection. Besides, we pioneered a hybrid reasoning model that supports both fast thinking and deep thinking within a unified framework to enhance the reasoning capabilities, diverging from the conventional industry practice of deploying two separate models. Through extensive experimental evaluations, our model demonstrates state-of-the-art SEA multilingual and e-commerce performance among sub-30B models, while maintaining significantly lower inference cost.

The recent emergence of Large Language Models based on the Transformer architecture has enabled dramatic advancements in the field of Natural Language Processing. However, these models have long inference latency, which limits their deployment and makes them prohibitively expensive for various real-time applications. The inference latency is further exacerbated by autoregressive generative tasks, as models need to run iteratively to generate tokens sequentially without leveraging token-level parallelization. To address this, we propose Big Little Decoder (BiLD), a framework that can improve inference efficiency and latency for a wide range of text generation applications. The BiLD framework contains two models with different sizes that collaboratively generate text. The small model runs autoregressively to generate text with a low inference cost, and the large model is only invoked occasionally to refine the small model's inaccurate predictions in a non-autoregressive manner. To coordinate the small and large models, BiLD introduces two simple yet effective policies: (1) the fallback policy that determines when to hand control over to the large model; and (2) the rollback policy that determines when the large model needs to correct the small model's inaccurate predictions. To evaluate our framework across different tasks and models, we apply BiLD to various text generation scenarios encompassing machine translation on IWSLT 2017 De-En and WMT 2014 De-En, and summarization on XSUM and CNN/DailyMail. On an NVIDIA T4 GPU, our framework achieves a speedup of up to 2.12x speedup with minimal generation quality degradation. Furthermore, our framework is fully plug-and-play and can be applied without any modifications in the training process or model architecture. Our code is open-sourced

Deploying transformer models in practice is challenging due to their inference cost, which scales quadratically with input sequence length. To address this, we present a novel Learned Token Pruning (LTP) method which adaptively removes unimportant tokens as an input sequence passes through transformer layers. In particular, LTP prunes tokens with an attention score below a threshold value which is learned for each layer during training. Our threshold-based method allows the length of the pruned sequence to vary adaptively based on the input sequence, and avoids algorithmically expensive operations such as top-k token selection. We extensively test the performance of LTP on GLUE tasks and show that our method outperforms the prior state-of-the-art token pruning methods by up to ~2.5% higher accuracy with the same amount of FLOPs. In particular, LTP achieves up to 2.1x FLOPs reduction with less than 1% accuracy drop, which results in up to 1.9x and 2.0x throughput improvement on Intel Haswell CPUs and NVIDIA V100 GPUs, respectively. Furthermore, we demonstrate that LTP is more robust than prior methods to variations on input sentence lengths. Our code has been developed in PyTorch and has been open-sourced.

We present a case study where an automatic AI system formalizes a textbook with more than 500 pages of graduate-level algebraic combinatorics to Lean. The resulting formalization represents a new milestone in textbook formalization scale and proficiency, moving from early results in undergraduate topology and restructuring of existing library content to a full standalone formalization of a graduate textbook. The formalization comprises 130K lines of code and 5900 Lean declarations and was conducted within one week by a total of 30K Claude 4.5 Opus agents collaborating in parallel on a shared code base via version control, simultaneously setting a record in multi-agent software engineering with usable results. The inference cost matches or undercuts what we estimate as the salaries required for a team of human experts, and we expect there is still the potential for large efficiencies to be made without the need for better models. We make our code, the resulting Lean code base and a side-by-side blueprint website available open-source.

We show that domain-general automatic evaluators can significantly improve the performance of agents for web navigation and device control. We experiment with multiple evaluation models that trade off between inference cost, modularity of design, and accuracy. We validate the performance of these models in several popular benchmarks for digital agents, finding between 74.4 and 92.9% agreement with oracle evaluation metrics. Finally, we use these evaluators to improve the performance of existing agents via fine-tuning and inference-time guidance. Without any additional supervision, we improve state-of-the-art performance by 29% on the popular benchmark WebArena, and achieve a 75% relative improvement in a challenging domain transfer scenario.

Diffusion- and flow-based generative models have recently demonstrated strong performance in protein backbone generation tasks, offering unprecedented capabilities for de novo protein design. However, while achieving notable performance in generation quality, these models are limited by their generating speed, often requiring hundreds of iterative steps in the reverse-diffusion process. This computational bottleneck limits their practical utility in large-scale protein discovery, where thousands to millions of candidate structures are needed. To address this challenge, we explore the techniques of score distillation, which has shown great success in reducing the number of sampling steps in the vision domain while maintaining high generation quality. However, a straightforward adaptation of these methods results in unacceptably low designability. Through extensive study, we have identified how to appropriately adapt Score identity Distillation (SiD), a state-of-the-art score distillation strategy, to train few-step protein backbone generators which significantly reduce sampling time, while maintaining comparable performance to their pretrained teacher model. In particular, multistep generation combined with inference time noise modulation is key to the success. We demonstrate that our distilled few-step generators achieve more than a 20-fold improvement in sampling speed, while achieving similar levels of designability, diversity, and novelty as the Proteina teacher model. This reduction in inference cost enables large-scale in silico protein design, thereby bringing diffusion-based models closer to real-world protein engineering applications. The PyTorch implementation is available at https://github.com/LY-Xie/SiD_Protein

Large language models (LLMs) have demonstrated impressive abilities in various domains while the inference cost is expensive. The state-of-the-art methods use 2-bit quantization for mainstream LLMs. However, challenges still exist: (1) Nonnegligible accuracy loss for 2-bit quantization. Weights are quantized by groups, while the ranges of weights are large in some groups, resulting in large quantization errors and nonnegligible accuracy loss (e.g. >3% for Llama2-7b with 2-bit quantization in GPTQ and Greenbit). (2) Limited accuracy improvement by adding 4-bit weights. Increasing 10% extra average bit more 4-bit weights only leads to <0.5% accuracy improvement on a quantized Llama2-7b. (3) Time-consuming dequantization operations on GPUs. The dequantization operations lead to >50% execution time, hindering the potential of reducing LLM inference cost. To tackle these challenges, we propose the following techniques: (1) We only quantize a small fraction of groups with the larger range using 4-bit with memory alignment consideration on GPUs.(2) We design the asynchronous dequantization on GPUs, leading to up to 3.92X speedup. We conduct extensive experiments on different model sizes. We achieve 2.85-bit for each weight and the end-to-end speedup for Llama2-7b is 1.74X over the original model, and we reduce both runtime cost and hardware cost by up to 2.70X and 2.81X with less GPU requirements.

We introduce SigLIP 2, a family of new multilingual vision-language encoders that build on the success of the original SigLIP. In this second iteration, we extend the original image-text training objective with several prior, independently developed techniques into a unified recipe -- this includes captioning-based pretraining, self-supervised losses (self-distillation, masked prediction) and online data curation. With these changes, SigLIP 2 models outperform their SigLIP counterparts at all model scales in core capabilities, including zero-shot classification, image-text retrieval, and transfer performance when extracting visual representations for Vision-Language Models (VLMs). Furthermore, the new training recipe leads to significant improvements on localization and dense prediction tasks. We also train variants which support multiple resolutions and preserve the input's native aspect ratio. Finally, we train on a more diverse data-mixture that includes de-biasing techniques, leading to much better multilingual understanding and improved fairness. To allow users to trade off inference cost with performance, we release model checkpoints at four sizes: ViT-B (86M), L (303M), So400m (400M), and g (1B).

Large language models have recently enabled a generative paradigm for query expansion, but their high inference cost makes direct deployment difficult in practical retrieval systems. To address this issue, a retrieval-feedback-driven distillation and preference-alignment framework is proposed to transfer retrieval-friendly expansion behavior from a strong teacher model to a compact student model. Rather than relying on few-shot exemplars at inference time, the framework first leverages two complementary types of teacher-generated expansions, produced under zero-shot and few-shot prompting conditions, as supervision signals for distillation and as candidate pools for preference construction. A retrieval-metric-driven strategy is then introduced to automatically form chosen/rejected expansion pairs according to nDCG@10 differences, and Direct Preference Optimization is applied to explicitly align generation preferences with retrieval objectives. Experiments on TREC DL19/20/21 and MIRACL-zh show that the proposed approach preserves strong retrieval effectiveness while substantially reducing inference cost. In particular, the distilled Qwen3-4B model reaches about 97% of the teacher (DeepSeek-685B) model's nDCG@10 performance on DL19, and remains effective on the Chinese MIRACL-zh benchmark, demonstrating strong practicality across both English and Chinese retrieval settings.

Large language models (LLMs) often seamlessly adapt to new tasks through in-context learning (ICL) or supervised fine-tuning (SFT). However, ICL is inefficient when handling many demonstrations, and SFT incurs training overhead while sacrificing flexibility. Mapping instructions or demonstrations from context directly into adapter parameters offers an appealing alternative. While prior work explored generating adapters based on a single input context, it has overlooked the need to integrate multiple chunks of information. To address this gap, we introduce CompAs, a meta-learning framework that translates context into adapter parameters with a compositional structure that allows them to be merged algebraically. This approach yields three benefits: lower inference cost, improved stability under long contexts, and establishes a principled solution when input exceeds the model's context window. Furthermore, CompAs reversibly encodes information into adapter parameters, enabling recovery of the original input context and facilitating safety. Empirical results on diverse multiple-choice and extractive question answering tasks show that CompAs outperforms ICL and prior generator-based methods, especially when scaling to more inputs. Our work establishes composable adapter generation as a practical and efficient alternative for scaling LLM deployment.

Recent advancements in large language models (LLMs) have significantly advanced complex reasoning capabilities, particularly through extended chain-of-thought (CoT) reasoning that incorporates mechanisms such as backtracking, self-reflection and self-correction. Despite these developments, the self-correction abilities of LLMs during long CoT reasoning remain underexplored. And recent findings on overthinking suggest that such models often engage in unnecessarily redundant reasoning. In this work, we empirically show that the first reasoning step exerts a disproportionately large influence on the final prediction - errors introduced at this stage can substantially degrade subsequent reasoning quality. This phenomenon is consistently observed across two state-of-the-art open-source reasoning model families: DeepSeek-R1 and Qwen3. To address this, we propose an efficient sampling strategy that leverages a reward model to identify and retain high-quality first reasoning steps while discarding suboptimal ones, achieving up to a 70% reduction in inference cost without sacrificing accuracy. Finally, we introduce a new benchmark specifically constructed with deliberately flawed first reasoning steps to systematically evaluate model self-correction capabilities, offering a foundation for future research on robust reasoning in LLMs.

Diffusion models have achieved remarkable generation quality, but they suffer from significant inference cost due to their reliance on multiple sequential denoising steps, motivating recent efforts to distill this inference process into a few-step regime. However, existing distillation methods typically approximate the teacher trajectory by using linear shortcuts, which makes it difficult to match its constantly changing tangent directions as velocities evolve across timesteps, thereby leading to quality degradation. To address this limitation, we propose ArcFlow, a few-step distillation framework that explicitly employs non-linear flow trajectories to approximate pre-trained teacher trajectories. Concretely, ArcFlow parameterizes the velocity field underlying the inference trajectory as a mixture of continuous momentum processes. This enables ArcFlow to capture velocity evolution and extrapolate coherent velocities to form a continuous non-linear trajectory within each denoising step. Importantly, this parameterization admits an analytical integration of this non-linear trajectory, which circumvents numerical discretization errors and results in high-precision approximation of the teacher trajectory. To train this parameterization into a few-step generator, we implement ArcFlow via trajectory distillation on pre-trained teacher models using lightweight adapters. This strategy ensures fast, stable convergence while preserving generative diversity and quality. Built on large-scale models (Qwen-Image-20B and FLUX.1-dev), ArcFlow only fine-tunes on less than 5% of original parameters and achieves a 40x speedup with 2 NFEs over the original multi-step teachers without significant quality degradation. Experiments on benchmarks show the effectiveness of ArcFlow both qualitatively and quantitatively.

Deployment of Transformer models on edge devices is becoming increasingly challenging due to the exponentially growing inference cost that scales quadratically with the number of tokens in the input sequence. Token pruning is an emerging solution to address this challenge due to its ease of deployment on various Transformer backbones. However, most token pruning methods require computationally expensive fine-tuning, which is undesirable in many edge deployment cases. In this work, we propose Zero-TPrune, the first zero-shot method that considers both the importance and similarity of tokens in performing token pruning. It leverages the attention graph of pre-trained Transformer models to produce an importance distribution for tokens via our proposed Weighted Page Rank (WPR) algorithm. This distribution further guides token partitioning for efficient similarity-based pruning. Due to the elimination of the fine-tuning overhead, Zero-TPrune can prune large models at negligible computational cost, switch between different pruning configurations at no computational cost, and perform hyperparameter tuning efficiently. We evaluate the performance of Zero-TPrune on vision tasks by applying it to various vision Transformer backbones and testing them on ImageNet. Without any fine-tuning, Zero-TPrune reduces the FLOPs cost of DeiT-S by 34.7\% and improves its throughput by 45.3\% with only 0.4\% accuracy loss. Compared with state-of-the-art pruning methods that require fine-tuning, Zero-TPrune not only eliminates the need for fine-tuning after pruning but also does so with only 0.1\% accuracy loss. Compared with state-of-the-art fine-tuning-free pruning methods, Zero-TPrune reduces accuracy loss by up to 49\% with the same or higher throughput.

The Mixture of Experts (MoE) architecture reduces the training and inference cost significantly compared to a dense model of equivalent capacity. Upcycling is an approach that initializes and trains an MoE model using a pre-trained dense model. While upcycling leads to initial performance gains, the training progresses slower than when trained from scratch, leading to suboptimal performance in the long term. We propose Drop-Upcycling - a method that effectively addresses this problem. Drop-Upcycling combines two seemingly contradictory approaches: utilizing the knowledge of pre-trained dense models while statistically re-initializing some parts of the weights. This approach strategically promotes expert specialization, significantly enhancing the MoE model's efficiency in knowledge acquisition. Extensive large-scale experiments demonstrate that Drop-Upcycling significantly outperforms previous MoE construction methods in the long term, specifically when training on hundreds of billions of tokens or more. As a result, our MoE model with 5.9B active parameters achieves comparable performance to a 13B dense model in the same model family, while requiring approximately 1/4 of the training FLOPs. All experimental resources, including source code, training data, model checkpoints and logs, are publicly available to promote reproducibility and future research on MoE.