ICML 2026 论文列表

Speculative decoding has emerged as a widely adopted paradigm for accelerating large language model inference, where a lightweight draft model rapidly generates candidate tokens that are then verified in parallel by a larger target model. However, due to limited model capacity, drafts often struggle to approximate the target distribution, resulting in shorter acceptance lengths and diminished speedup. A key yet under-explored observation is that speculative decoding inherently provides *verification feedback* that quantifies the deviation between the draft and target models at no additional cost. This process naturally forms an iterative "draft commits--feedback provides--draft adapts" evolving loop, which precisely matches the *online learning* paradigm. Motivated by this connection, we propose OnlineSPEC, a unified framework that systematically leverages interactive feedback to continuously evolve draft models. Grounded in *dynamic regret minimization*, we establish a formal link between online learning performance and speculative system's acceleration rate, and develop novel algorithms via modern online learning techniques, including optimistic online learning that adaptively reuses historical gradients as predictive update hints, and online ensemble learning that dynamically maintains multiple draft models. Our algorithms are equipped with theoretical justifications and improved acceleration rates, achieving up to 24% speedup over seven benchmarks and five foundation models.

Multimodal large language models (MLLMs) struggle with numerical regression under longtailed target distributions. Token-level supervised fine-tuning (SFT) and point-wise regression rewards bias learning toward high-density regions, leading to regression-to-the-mean behavior and poor tail performance. We identify the lack of cross-sample relational supervision as a key limitation of existing MLLM training paradigms. To address it, we propose a distribution-aware reinforcement learning framework based on Group Relative Policy Optimization, which introduces batch-level comparison-based supervision via the Concordance Correlation Coefficient-based reward to align predicted and ground-truth distributions in terms of correlation, scale, and mean. The framework is plug-and-play, requiring no architectural modification. Experiments on a unified suite of long-tailed regression benchmarks show consistent improvements over SFT and existing MLLM regression methods, with particularly strong gains in medium- and few-shot regimes.

Despite rapid progress in Grounded Visual Reasoning (GVR) with MLLMs and RL-style fine-tuning, existing approaches often lack effective learning signals for intermediate grounding decisions and are prone to shortcut solutions. In this work, we explicitly decompose GVR into Evidence Generation followed by Counterfactual Answer Reasoning, and formalize this structure as a Causal Grounding Graph (CGG) in which the generated evidence acts as a causal mediator. Building on this formulation, we propose CofactGVR, which estimates the mediator’s utility via a matched counterfactual intervention that perturbs the predicted region while keeping the original image–question context fixed. The factual–counterfactual reward gap yields a principled intermediate bonus, selectively assigned to high-quality factual rollouts to promote evidence-faithful reasoning. To further stabilize and efficiently exploit this causal training signal, we incorporate a Quantile-filtered Prioritized Advantage Sampling (Q-PAS) strategy that preferentially updates on trajectories with high-magnitude advantages while filtering low-signal samples. Extensive experiments across GVR benchmarks show consistent improvements, indicating that CofactGVR strengthens reliance on informative visual evidence under controlled interventions. Our source code will be released at https://github.com/zhangy0822/CofactGVR.

Manufacturable chip layouts must satisfy thousands of geometry-based design rules, and design rule checking (DRC) enforces them by running executable DRC scripts on layouts. Translating natural language rules into correct DRC scripts is labor-intensive and requires specialized expertise, motivating LLM agents for DRC script synthesis and debugging. However, existing benchmarks have small evaluation sets and often evaluate scripts by code similarity rather than execution correctness, and prior machine learning-based methods either ignore execution feedback or require labeled test layouts as agent's input. To this end, we introduce Rule2DRC, a large-scale benchmark for DRC script coding agents with 1,000 rule-to-script tasks and 13,921 evaluation chip layouts for execution-based scoring. Rule2DRC provides an evaluation pipeline that measures functional correctness via DRC execution outcomes without requiring evaluation layouts as input to the agent. We also propose SplitTester, a tester agent for program selection that uses execution feedback to generate discriminative test cases and separate previously indistinguishable candidate scripts, substantially improving Best-of-N selection performance in this domain. We release the code at https://github.com/snu-mllab/Rule2DRC.

4-bit quantization reduces the memory footprint and latency of large language model inference, but its aggressive precision reduction can severely degrade accuracy. Prior methods address this by decomposing each weight matrix into two components (e.g., via singular value decomposition) and quantizing them separately, assigning the bulk of values to a low-precision residual component while handling outliers with a high-precision low-rank component. However, such decompositions are designed to minimize the real-valued energy of the residual, rather than the post-quantization error of the residual and low-rank components. We propose TwinQuant, a 4-bit quantization framework that learns quantization-friendly decomposed subspaces and jointly reshapes both the low-rank and residual components. TwinQuant learns component-specific transformations via a joint optimization over the Stiefel and general linear manifolds, flattening their distributions and reducing dynamic-range imbalance. To enable efficient end-to-end execution, we further design a fused dual-component kernel that pipelines the two-stage low-rank computation on-chip and merges both components with a single epilogue, avoiding intermediate global-memory traffic. Across LLaMA3 and Qwen3 models, TwinQuant preserves near-FP16 accuracy and delivers up to $1.8\times$ end-to-end speedup over an FP16 baseline.

Large language models have demonstrated remarkable performance; however, their massive parameter counts make deployment highly expensive. Low-rank approximation offers a promising compression solution, yet existing approaches have two main limitations: (1) They focus on minimizing the output error of individual linear layers, without considering the architectural characteristics of Transformers, and (2) they decompose a large weight matrix into two small low-rank matrices. Consequently, these methods often fall short compared to other compression techniques like pruning and quantization, and introduce runtime overhead such as the extra GEMM kernel launches and memory operations for decomposed small matrices. To address these limitations, we propose $A^3$, a post-training low-rank approximation framework. $A^3$ splits a Transformer layer into three functional components, namely $\texttt{QK}$, $\texttt{OV}$, and $\texttt{MLP}$ and provides analytical solutions that reduces the hidden dimension size inside each component while minimizing the component's functional loss. This approach directly reduces model sizes, KV cache sizes, and FLOPs without introducing any runtime overheads. Through extensive experiments, we show that $A^3$ maintains superior performance compared to SoTAs. For example, under the same reduction budget in computation and memory, our low-rank approximated LLaMA 3.1-70B achieves a perplexity of 4.69 on WikiText-2, outperforming the previous SoTA's 7.87 by 3.18. We also show versatile applications of $A^3$ in KV cache compression, integration with quantization, fine-tuning and mixed-rank assignments. We open-sourced our framework at https://github.com/DeepWok/a3.

Aligning large language models (LLMs) to heterogeneous and rapidly evolving safety requirements remains a critical challenge. Existing instruction-tuned LLMs and standalone safety classifiers often fail to generalize to new safety configurations, motivating the need for Reward Models (RMs) that are explicitly configurable to changing specifications. We introduce the Configurable Safety Reward Model (CSRM), which is jointly optimized for calibrated safety compliance and reward modeling. Our approach is supported by configuration-targeted data augmentation that enforces instruction adherence while preserving relative severity structure. The resulting RM is sensitive to fine-grained safety configurations and conversational nuances, substantially improving generalization to previously unseen safety configurations. CSRM achieves state-of-the-art performance on recent configurable safety benchmarks, including CoSApien (94.6\% F1) and DynaBench (75.8\% F1), without requiring additional human annotation. When used for downstream safety alignment, CSRM yields LLMs with a significantly improved helpfulness–safety tradeoff compared to existing baselines.

Omni-modal Large Language Models (Omni-LLMs) have demonstrated strong capabilities in audio-video understanding tasks. However, their reliance on long multimodal token sequences leads to substantial computational overhead. Despite this challenge, token compression methods designed for Omni-LLMs remain limited. To bridge this gap, we propose OmniSIFT (Omni-modal Spatio-temporal Informed Fine-grained Token compression), a modality-asymmetric token compression framework tailored for Omni-LLMs. Specifically, OmniSIFT adopts a two-stage compression strategy: (i) a spatio-temporal video pruning module that removes video redundancy arising from both intra-frame structure and inter-frame overlap, and (ii) a vision-guided audio selection module that filters audio tokens. The entire framework is optimized end-to-end via a differentiable straight-through estimator. Extensive experiments on five representative benchmarks verify the efficacy and robustness of OmniSIFT. Notably, for Qwen2.5-Omni-7B, OmniSIFT adds 4.85M parameters while still achieving lower latency than training-free baselines such as OmniZip. With only 25% of the original token context, OmniSIFT consistently outperforms all compression baselines and even surpasses the full-token model on several tasks.

Large Reasoning Models (LRMs) achieve strong problem-solving through long chain-of-thought, but their deployment is constrained by the high cost of full-precision inference and growing KV cache footprints. Microscaled FP4 formats enable efficient FP4 deployment; however, fully quantizing weights, activations, and KV caches (W4A4KV4) causes severe reasoning degradation that existing PTQ and QAT fail to recover. We identify that FP4 failures concentrate on low-entropy tokens—precise symbolic commitments such as digits and operators—where quantization noise inflates sampling errors that cascade through reasoning traces. Based on this insight, we propose ReQAT, a reasoning-centric FP4 training framework with three components: (i) Trace-Aligned QAT (TAQ), which revisits identical reasoning traces to focus updates on critical low-entropy decisions; (ii) Selective Entropy Minimization (SEM), which reinforces confidence at low-entropy positions; and (iii) Q-FIT, a quantization-friendly initialization that jointly calibrates RoPE-consistent KV cache transformations to stabilize QAT. Under the same training budget, ReQAT not only recovers but surpasses BF16 fine-tuning accuracy, while delivering up to $3.9\times$ throughput speedup on NVIDIA DGX Spark and $3.1\times$ on B200. The project repository is available at https://github.com/aiha-lab/ReQAT.

Transformer attention is typically implemented using softmax normalization, which enforces attention weights with unit sum normalization. While effective in many settings, this constraint can limit flexibility in controlling attention magnitudes and may contribute to overly concentrated or unstable attention patterns during training. Prior work has explored modifications such as attention sinks or gating mechanisms, but these approaches provide only limited or indirect control over attention reweighting. We propose Affine-Scaled Attention, a simple extension to standard attention that introduces input-dependent scaling and a corresponding bias term applied to softmax-normalized attention weights. This design relaxes the strict normalization constraint while maintaining aggregation of value representations, allowing the model to adjust both the relative distribution and the scale of attention in a controlled manner. We empirically evaluate Affine-Scaled Attention in large-scale language model pretraining across multiple model sizes. Experimental results show consistent improvements in training stability, optimization behavior, and downstream task performance compared to standard softmax attention and attention sink baselines. These findings suggest that modest reweighting of attention outputs provides a practical and effective way to improve attention behavior in Transformer models.

We build \textbf{AICrypto}, a comprehensive benchmark designed to evaluate the cryptography capabilities of large language models (LLMs). The benchmark comprises 135 multiple-choice questions, 150 capture-the-flag challenges, and 30 proof problems, covering a broad range of skills from knowledge memorization to vulnerability exploitation and formal reasoning. All tasks are carefully reviewed or constructed by cryptography experts to improve correctness and rigor. For each proof problem, we provide detailed scoring rubrics and reference solutions that enable automated grading, achieving high correlation with human expert evaluations. We introduce strong human expert performance baselines for comparison across all task types. Our evaluation of 17 leading LLMs reveals that state-of-the-art models match or even surpass human experts in memorizing cryptographic concepts, exploiting common vulnerabilities, and routine proofs. However, our analysis reveals that they still lack a deep understanding of abstract mathematical concepts and struggle with tasks that require multi-step reasoning and dynamic analysis. We hope this work could provide insights for future research on LLMs in cryptographic applications. Our code and dataset are available at https://github.com/wangyu-ovo/aicrypto-agent.

Black box auditing of language models is an essential pre-deployment tool, but it may miss subtle forms of misalignment and hidden information. To better elicit hidden information during an auditing process, we introduce *overthinking*: the process of using reasoning task vectors to amplify the chain-of-thought faithfulness of reasoning models. Given the parameters of a base instruct model M and reasoning-distilled model R, we define the *overthinking model* as $\mathcal{O}_\alpha = M + \alpha(R - M)$, where $\alpha > 1$ amplifies reasoning beyond the pure reasoning model R. Additionally, we introduce new layer-wise attenuation strategies that selectively amplify reasoning without losing quality and coherence of model outputs. We demonstrate that overthinking models are more likely to reveal hidden information across four experimental settings, across 2B-32B models. Our findings suggest that reasoning amplification may surface secrets or unintended behaviors acquired during training up to $10\times$ more frequently than the original reasoning model. How secrets surface depends on the secret type: some require perturbation along the reasoning direction, while others yield to any sufficiently large weight perturbation.

As the demand for emotional intelligence in large language models (LLMs) grows, a key challenge lies in understanding the internal mechanisms that give rise to emotional expression and in controlling emotions in generated text. This study addresses three core questions: (1) Do LLMs contain context-agnostic mechanisms shaping emotional expression? (2) What form do these mechanisms take? (3) Can they be harnessed for universal emotion control? We first construct a controlled dataset, $\textit{SEV}$ (Scenario–Event with Valence), to elicit comparable internal states across emotions. Subsequently, we extract context-agnostic emotion directions that reveal consistent, cross-context encoding of emotion (Q1). We identify neurons and attention heads that locally implement emotional computation through analytical decomposition and causal analysis, and validate their causal roles via ablation and enhancement interventions. Next, we quantify each sublayer's causal influence on the model’s final emotion representation and integrate the identified local components into coherent global emotion circuits that drive emotional expression (Q2). Directly modulating these circuits achieves 99.65% emotion-expression accuracy on the test set, surpassing prompting- and steering-based methods (Q3). To our knowledge, this is the first systematic study to uncover and validate emotion circuits in LLMs, offering new insights into interpretability and controllable emotional intelligence.

While the emerging field of agentic software engineering has spurred extensive research into post-training, this paradigm alone does not fully address the distribution mismatch between traditional static pre-training and dynamic deployment environments. In this paper, we instead investigate agentic mid-training as a scalable complementary approach. Central to our approach is *agent-native data* comprising two complementary components: *contextually-native trajectories* that preserve the complete information flow an agent experiences, offering broad coverage and diversity; and *environmentally-native trajectories* whose observations stem from actual tool invocations and test executions, providing interaction authenticity. On `SWE-Bench Verified`, our recipe outperforms the previous open software engineering mid-training recipe `Kimi-Dev` under two post-training settings with the same base model and agentic scaffold, while using fewer than half mid-training tokens (73.1B). Furthermore, our 32B and 72B models achieve state-of-the-art resolution rates of **56.1\%** and **58.5\%** among open agentic recipes using agentic scaffolds, despite starting from non-coder `Qwen2.5` base models. We also observe performance gains on general code generation and scientific benchmarks. We open-source a significant portion of our datasets, recipes, and model checkpoints to facilitate further research.

Conventional LLMs may suffer from corpus heterogeneity and subtle changes in conditions. While finetuning can create the catastrophe forgetting issue, applications of meta-learning on LLMs are also limited due to their complexity and scalability. In this paper, we activate the meta-signal of $\beta$ within the SwiGLU blocks, resulting in a meta-gating mechanism that adaptively adjusts the nonlinearity of FFN. A hypernetwork is employed to dynamically produce $\beta$ under textual conditions, providing meta-controllability over LLMs. By testing on different condition types such as task, domain, persona, and style, our method outperforms finetuning and meta-learning baselines, and can generalize reasonably on unseen tasks, condition types, or instructions. Our codes are in https://github.com/AaronJi/MeGan.

Typical reinforcement learning (RL) methods for LLM reasoning waste compute on hard problems, where correct on-policy traces are rare and policy gradients vanish. To bootstrap more efficient RL, we consider reusing old sampling FLOPs (from prior inference or RL training) in the form of off-policy traces. We introduce PrefixRL, where we condition on the prefix of successful off-policy traces and run on-policy RL to complete them, side-stepping instabilities from using off-policy data as supervision targets. PrefixRL boosts the learning signal on hard problems by modulating the difficulty of the problem through the off-policy prefix length. We prove that the PrefixRL objective is not only consistent with the standard RL objective but also more sample efficient. Empirically, we discover back-generalization: training only on prefixed problems generalizes to out-of-distribution unprefixed performance, with learned strategies often differing from those in the prefix. In our experiments, we source the off-policy traces by rejection sampling with the base model, creating a self-improvement loop. On hard reasoning problems, PrefixRL reaches the same training reward 2x faster than the strongest baseline (SFT on off-policy data then RL), even after accounting for the compute spent on the initial rejection sampling, and increases the final reward by 3x.

Masked diffusion language models enable parallel token generation and offer improved decoding efficiency over autoregressive models. However, their performance degrades significantly when generating multiple tokens simultaneously, due to a mismatch between token-level training objectives and joint sequence consistency. In this paper, we propose ME-DLM, an edit-based refinement framework that augments diffusion generation with lightweight post-editing steps. After producing an initial complete response, the model refines it through minimal edit operations, including replacement, deletion, and insertion, conditioned on the full sequence. Training supervision is derived from edit distance, providing a deterministic signal under a fixed canonicalization scheme for learning minimal corrections. This approach encourages sequence-level consistency through globally conditioned edits while preserving the efficiency benefits of parallel diffusion decoding. Extensive experiments demonstrate that ME-DLM improves the quality and robustness of multi-token parallel generation. In particular, when built upon LLaDA, our method achieves consistent gains of 11.6 points on HumanEval and 33.6 points on GSM8K while using one-eighth of the total diffusion steps. Code is available at https://github.com/renhouxing/ME-DLM.

In many applications of LLMs, natural language responses often have an underlying structure such as representing discrete labels, numerical values, or graphs. Yet, existing decoding and uncertainty estimation methods operate only in language space and largely disregard structural information. We address this by modeling LLM outputs directly in a task-dependent latent structure. By equipping this structure with a dissimilarity measure, we can compute Bayes-optimal responses. These are not selected from sampled generations but are newly synthesized by combining individual responses in the latent space. Across different tasks, Bayes-optimal responses consistently outperform standard decoding methods like beam search. Moreover, quantifying uncertainty via the induced Bayesian risk captures variations in terms of the latent structure and improves alignment with output quality and correctness. Our decision-theoretic framework is applicable to any problem that admits a latent response structure and enables reliable task-aware LLM predictions.

Quantizing LLM weights and activations is a standard approach for efficient deployment, but a few extreme outliers can stretch the dynamic range and amplify low-bit quantization errors. Prior transform-based mitigations (e.g., Hadamard rotations) are fixed and data-agnostic, and their optimality for quantization has remained unclear. We derive closed-form optimal linear blockwise transforms for joint weight-activation quantization under standard RTN AbsMax-scaled block quantizers, covering both integer and floating-point formats. The resulting construction, WUSH, combines a Hadamard backbone with a data-dependent second-moment component to form a non-orthogonal transform that is provably near-optimal for FP and INT quantizers under mild assumptions while admitting an efficient fused GPU implementation. Empirically, WUSH improves W4A4 accuracy over the strongest Hadamard-based baselines (e.g., on Llama-3.1-8B-Instruct in MXFP4, it gains +2.8 average points with RTN and +0.7 with GPTQ) while delivering up to 5.8$\times$ per-layer throughput over BF16 via FP4 MatMul. Source code is available at https://github.com/IST-DASLab/WUSH.