ICML 2026 Papers

Personalized federated LoRA fine tuning has become a key approach to addressing data heterogeneity in distributed fine tuning of large language models (LLMs). Existing methods typically assume homogeneous personalization needs across clients, relying on dual LoRA or personalized calibration schemes. However, they fail to account for the heterogeneity of local personalization requirements and the conflicting optimization objectives in dual LoRA, limiting scalability and performance. To address this, we propose Het-CPFLoRA, a customizable heterogeneous federated LoRA fine tuning algorithm inspired by the decoupling properties of LoRA parameters. We employ a single adapter fine tuning scheme to mitigate conflicts between personalized and generalized optimization, decouple LoRA into generalized and personalized subspaces for local customization, and use SVD compression to integrate cross client generalized knowledge. During inference, we introduce an OOD oriented dynamic mechanism to adjust the weighting between personalized and generalized decoupling knowledge, improving performance on user data. Extensive experiments on two public benchmark datasets show that Het-CPFLoRA outperforms state of the art methods in both personalization and generalization across heterogeneous scenarios. The code will be released as an open-source project.

Large Reasoning Models (LRMs) have recently achieved strong mathematical and code reasoning performance through Reinforcement Learning (RL) post-training. However, we show that modern reasoning post-training induces an unintended exploration collapse: temperature-based sampling no longer increases pass@$n$ accuracy. Empirically, the final-layer posterior of post-trained LRMs exhibit sharply reduced entropy, while the entropy of intermediate layers remains relatively high. Motivated by this entropy asymmetry, we propose Latent Exploration Decoding (LED), a depth-conditioned decoding strategy. LED aggregates intermediate posteriors via cumulative sum and selects depth configurations with maximal entropy as exploration candidates. Without additional training or parameters, LED consistently improves pass@1 and pass@16 accuracy by 0.61 and 1.03 percentage points across multiple reasoning benchmarks and models. Relevant code is included in the supplementary material and will made be fully public after this paper is accepted.

Agentic multimodal models have garnered significant attention for their ability to leverage external tools to tackle complex tasks. However, it is observed that such agents often meet premature interaction collapse, caused by two primary reasons: 1) the terminal reward often appending on the last token prevents the advantage from distinguishing trajectories with exploratory behavior; 2) excessively redundant context hinders the agent from absorbing useful feedback. To address these issues, we propose the Deepening Reasoning MMSearchAgent, the framework leverages the structural proximity to derive advantage signals from the whole rollout trajectories in an entire batch, such that trajectories of different lengths are further encouraged to be generated, even when containing the same correct answer. Additionally, differentiated gaussian rewards are employed to dynamically calibrate interaction tolerance, thereby ensuring information reliability and reducing redundancy. To support multi-turn interaction training, we have constructed a multi-step deep-reasoning dataset including 3602 high-quality QA pairs with at least 3 reasoning steps. Extensive experiments demonstrate that our method achieves state-of-the-art performance, outperforming the MMSearch-R1 by 8.4\% on FVQA-test.

Large Audio–Language Models (LALMs) are becoming essential as a powerful multimodal backbone for real-world applications. However, recent studies show that audio inputs can more easily elicit harmful responses than text, exposing new risks toward deployment. While safety alignment has made initial advances in LLMs and Large Vision–Language Models (LVLMs), we find that vanilla adaptation of these approaches to LALMs faces two key limitations: 1) LLM-based steering fails under audio input due to the large distributional gap between activations, and 2) prompt-based defenses induce over-refusals on benign-speech queries. To address these challenges, we propose **S**afe-**A**blated **R**efusal **Steer**ing (SARSteer), an effective inference-time defense framework for LALMs. Specifically, SARSteer leverages text-derived refusal steering to enforce rejection without manipulating audio inputs and introduces decomposed safe-space ablation to mitigate over-refusal. Extensive experiments demonstrate that SARSteer significantly improves harmful-query refusal while preserving benign responses, establishing a principled step toward safety alignment in LALMs. The codes and constructed datasets are released at https://github.com/linweiii/SARSteer.

Large language model (LLM)-driven multi-agent systems typically require multiple model invocations and complex coordination during inference, and their execution strategies directly affect system accuracy, latency, and computational cost. Parallel execution provides a means to improve inference-time efficiency. From the perspective of inference-time execution, this paper models parallelism in multi-agent systems as two distinct levels of decision processes: Replica Parallelism, which explores multiple complete solution paths at the task level, and Structural Parallelism, which enables concurrent execution within a single solution path through task decomposition. However, the roles of different forms of parallelism and their interrelationships still lack systematic study in terms of unified organization and coordination. We therefore propose TIPEX, a controllable execution framework that unifies these two levels of parallelism and coordinates their roles within the inference process under a unified execution semantics while supporting systematic combinations and analyses of different parallel strategies and parameter configurations. Systematic experiments on the GAIA benchmark demonstrate that inference-time parallelism can significantly improve accuracy and reduce end-to-end latency at the cost of increased token consumption. Further analysis shows that Replica and Structural Parallelism exhibit complementary effects across task complexities, with tasks of intermediate difficulty benefiting most from their coordination, while overly aggressive parallel strategies do not necessarily yield better performance.

Supervised fine-tuning (SFT) is the standard approach for post-training large language models (LLMs), yet it often shows limited generalization. We trace this limitation to its default training objective: negative log likelihood (NLL). While NLL is classically optimal when training from scratch, post-training operates in a different paradigm and could violate its optimality assumptions, where models already encode task-relevant priors and supervision can be long and noisy. In this work, we systematically study various probability-based objectives and characterize when and why different objectives succeed or fail under varying conditions. Through comprehensive experiments and extensive ablation studies across 8 model backbones, 27 benchmarks, and 7 domains, we uncover a critical dimension that governs objective behavior: the model-capability continuum. Near the model-strong end, prior-leaning objectives that downweight low-probability tokens (e.g., $-p$, $-p^{10}$, thresholded variants) consistently outperform NLL; toward the model-weak end, NLL dominates; in between, no single objective prevails. Our theoretical analysis further elucidates how objectives trade places across the continuum, providing a principled foundation for adapting objectives to model capability. The code is available at https://github.com/GaotangLi/Beyond-Log-Likelihood.

We study how syntactic and semantic information is encoded in inner layer representations of Large Language Models (LLMs), focusing on the very large DeepSeek-V3. We find that, by averaging hidden-representation vectors of sentences sharing syntactic structure or meaning, we obtain vectors that capture a significant proportion of the syntactic and semantic information contained in the representations. In particular, subtracting these syntactic and semantic ``centroids'' from sentence vectors strongly affects their similarity with syntactically and semantically matched sentences, respectively, suggesting that syntax and semantics are, at least partially, linearly encoded. We also find that the cross-layer encoding profiles of syntax and semantics are different, and that the two signals can to some extent be decoupled, suggesting differential encoding of these two types of linguistic information in LLM representations.

Evaluation of socially unsafe content in spoken dialogues remains text-centric, missing prosody and transcription failures. We present LALM-as-a-Judge, which includes an open benchmark of 24,000 multi-turn spoken dialogues with one localized unsafe turn, generated out of 8 socially unsafe categories and 5 severity levels. We evaluate 6 large audio-language models (LALMs) as judges, open and closed-source, in text-only, audio-only, and multimodal setups by their sensitivity, severity-order specificity, and turn-position bias for socially harmful content in the dialogue. Results show that audio contributes non-lexical evidence beyond transcript semantics and that multimodal gains are not universal but can be text-anchored, balanced, conservative, and interfering, which we link to the audio pathway bottlenecks and fusion limits. We position the benchmark as diagnostic and derive practitioner guidance for model, modality, and prompts choices.

Key–value (KV) caching is essential for large language model inference, yet its memory overhead poses a critical bottleneck for long-context generation. Existing eviction policies predominantly rely on empirical heuristics, lacking a rigorous theoretical foundation. This work rethinks KV cache eviction through the lens of the Information Bottleneck principle. Under a linear–Gaussian surrogate of attention, we derive a closed-form mutual information objective that characterizes the effective information capacity of a retained KV cache subset. This formulation reveals that a wide range of existing eviction strategies can be interpreted as different approximations of the same capacity-maximization principle. Guided by this insight, we introduce CapKV, a capacity-aware eviction method that directly targets information preservation via a log-determinant approximation using statistical leverage scores. This approach replaces heuristic selection with a theoretically grounded mechanism that preserves the maximum predictive signal. Extensive experiments across multiple models and long-context benchmarks show that CapKV consistently outperforms prior methods, achieving a better trade-off between memory efficiency and generational fidelity.

Current auto-regressive models can generate high-quality, topologically precise meshes; however, they necessitate thousands—or even tens of thousands—of next-token predictions during inference, resulting in substantial latency. We introduce XSpecMesh, a quality-preserving acceleration method for auto-regressive mesh generation models. XSpecMesh employs a lightweight, multi-head speculative decoding scheme to predict multiple tokens in parallel within a single forward pass, thereby accelerating inference. We further propose a verification and resampling strategy: the backbone model verifies each predicted token and resamples any tokens that do not meet the quality criteria. In addition, we propose a distillation strategy that trains the lightweight decoding heads by distilling from the backbone model, encouraging their prediction distributions to align and improving the success rate of speculative predictions. Extensive experiments demonstrate that our method achieves a $1.7\times$ speedup without sacrificing generation quality. Our code will be released.

Optimizing pretraining data composition is pivotal for LLM generalization. While dynamic mixing outperforms static strategies by capturing evolving training dynamics, current methods fail to reconcile computational efficiency with sample efficiency and structural flexibility for diverse pipelines.We introduce \textbf{Actor--Critic Online Data Mixing (AC-ODM)}, which approaches data mixing from a reinforcement learning perspective with a parameterized policy that we theoretically prove to act as a dynamic linear surrogate maximizing the constructive interference of gradients. To enhance practical flexibility, AC-ODM supports two operational modes: (i) a \textbf{proxy mode} for fixed, pre-prepared corpora, where a policy learned on a small model is transferred to a larger target; and (ii) a \textbf{non-proxy mode} for direct end-to-end training from scratch without priors. Empirically, AC-ODM significantly outperforms prior methods in convergence speed and downstream accuracy across various architectures. On Pythia-1B, it reaches optimal validation perplexity using up to 66\% fewer training steps than competitive baselines, delivering a 27.5\% relative improvement in MMLU accuracy and a 2.23$\times$ higher pass@1 on HumanEval, all while incurring a virtually negligible ($~$0.4\%) per-step wall-clock increase and only 2\% additional memory overhead. Code is available at \url{https://github.com/DANG-ai/AC-ODM}.

The design of Analog and Mixed-Signal (AMS) integrated circuits remains heavily reliant on expert knowledge, with transistor sizing a major bottleneck due to nonlinear behavior, high-dimensional design spaces, and strict performance constraints. Existing Electronic Design Automation (EDA) methods typically frame sizing as static black-box optimization, resulting in inefficient and less robust solutions. Although Large Language Models (LLMs) exhibit strong reasoning abilities, they are not suited for precise numerical optimization in AMS sizing. To address this gap, we propose \textsc{AutoSizer}, a reflective LLM-driven meta-optimization framework that unifies circuit understanding, adaptive search-space construction, and optimization orchestration in a closed loop. It employs a two-loop optimization framework, with an inner loop for circuit sizing and an outer loop that analyzes optimization dynamics and constraints to iteratively refine the search space from simulation feedback. We further introduce \textsc{AMS-SizingBench}, an open benchmark comprising 24 diverse AMS circuits in SKY130 CMOS technology, designed to evaluate adaptive optimization policies under realistic simulator-based constraints. \textsc{AutoSizer} experimentally achieves higher solution quality, faster convergence, and higher success rate across varying circuit difficulties, outperforming both traditional optimization methods and existing LLM-based agents.

Federated LoRA provides a communication-efficient mechanism for fine-tuning large language models on decentralized data. In practice, however, a discrepancy between the factor-wise averaging used to preserve low rank and the mathematically correct aggregation of local updates can cause significant aggregation error and unstable training. We argue that a major source of this problem is *rotational misalignment*, arising from the rotational invariance of low-rank factorizations -- semantically equivalent updates can be represented in different latent subspaces across clients since $(B_i R_i)(R_i^\top A_i) = B_i A_i$. When such misaligned factors are averaged directly, they interfere destructively and degrade the global update. To address this issue, we propose **FedRot-LoRA**, a federated LoRA framework that aligns client updates via orthogonal transformations prior to aggregation. This alignment preserves the semantic update while reducing cross-client subspace mismatch, without increasing communication cost or restricting model expressivity. We provide a convergence analysis that examines the aggregation error induced by factor-wise averaging and shows how rotational alignment yields a tighter upper bound on this error. Extensive experiments on natural language understanding and generative tasks demonstrate that FedRot-LoRA consistently outperforms existing federated LoRA baselines across a range of heterogeneity levels and LoRA ranks. The code is available at https://github.com/haoran-zh/FedRot-LoRA.

Data science agents promise to accelerate discovery and insight-generation by turning data into executable analyses and findings. Yet existing data science benchmarks fall short due to fragmented evaluation interfaces that make cross-benchmark comparison difficult, narrow task coverage and a lack of rigorous data grounding. In particular, we show that a substantial portion of tasks in current benchmarks can be solved without using the actual data. To address these limitations, we introduce DSGym, a standardized framework for evaluating and training data science agents in self-contained execution environments. Unlike static benchmarks, DSGym provides a modular architecture that makes it easy to add tasks, agent scaffolds, and tools, positioning it as a live, extensible testbed. We curate DSGym-Tasks, a holistic task suite that standardizes and refines existing benchmarks via quality and shortcut solvability filtering. We further expand coverage with (1) DSBio: expert-derived bioinformatics tasks grounded in literature and (2) DSPredict: challenging prediction tasks spanning domains such as computer vision, molecular prediction, and single-cell perturbation. Beyond evaluation, DSGym enables agent training via execution-verified data synthesis pipeline. As a case study, we build a 2,000-example training set in DSGym that substantially improves a 4B mode on standardized analysis benchmarks. Overall, DSGym enables rigorous end-to-end measurement of whether agents can plan, implement, and validate data analyses in realistic scientific context.

Best-of-N selection improves reasoning in large language models (LLMs) by allocating test-time compute to sample candidate trajectories, but it relies on reliable verification. Widely used proxies have complementary failure modes: logit-confidence signals can suffer from calibration collapse, where confidence becomes misaligned with correctness, while sample agreement can be costly or brittle. Instead, we analyze the model's latent dynamics during inference. Motivated by metastable dynamics in cognitive systems, we introduce Latent Velocity Entropy (LVE), a training-free metric that quantifies the temporal concentration of internal representation updates. Experiments on four reasoning benchmarks, AIME25, GPQA, MATH500, and BRUMO25, show that LVE-based selection remains informative when logit-confidence baselines fail, with the clearest gains on reasoning-dense mathematical tasks and more modest gains on knowledge-heavy GPQA. On MATH500, LogNorm-LVE reaches 92.2\% Pass@1 at $N=10$ and nearly matches 10-sample majority voting with only 3 samples. Cross-model, bootstrap, and length-controlled analyses indicate that LVE is not reducible to sequence length, while also clarifying the task-dependent limits of the signal. Code is available at https://github.com/ZopenDP/LatentVelocityEntropy.

Recent research has made growing efforts to leverage large language models (LLMs) for computer-aided design (CAD), a domain that demands advanced geometric and spatial reasoning across long operation sequence. However, existing studies remain limited in addressing complex modeling tasks that necessitate step-by-step reasoning, primarily due to the scarcity of high-quality CAD datasets and the absence of fine-grained evaluation frameworks. In response to these challenges, we introduce Op-CAD, the first large-scale, multi-modal dataset for operation-oriented CAD generation, encompassing four operation types and five modalities. Furthermore, we introduce a novel CAD parsing module together with a geometry-guided hierarchical annotation pipeline, which decomposes modeling sequences into discrete operations and substantially improves the annotation accuracy of Vision-Language Models (VLMs). Based on our dataset, we redefine the CAD modeling task by decoupling geometric and spatial perspectives and introduce a novel metric, Chamfer/Fillet Intersection over Union (CF-IoU), to fill the void in assessing chamfer and fillet operations. By comprehensively evaluating eight LLMs on Op-CAD, we establish a benchmark for current models on operation-oriented tasks. Finally, we investigate performance enhancement strategies through fine-tuning on Op-CAD and propose Chain-of-Operation (COOP), a novel prompting strategy that emulates human-engineer reasoning.

Post-training with explicit reasoning traces is common to improve the reasoning capabilities of Multimodal Large Language Models (MLLMs). However, acquiring high-quality reasoning traces is often costly and time-consuming. Hence, the self-improvement paradigm has emerged, enabling MLLMs to self-generate reasoning traces for training without external supervision. Despite its effectiveness, we reveal two shortcomings in the self-improvement training of MLLMs: 1) data imbalance, where simple samples are over-trained, but the challenging yet crucial samples are under-trained; 2) language prior bias, where MLLMs overly rely on linguistic priors while neglecting the visual cues. To this end, we propose VISTA, a VIsion-aware Self-improvement Training framework for enhancing the multimodal Reasoning of MLLMs. Specifically, VISTA first introduces a prefix resampling strategy to reuse the partial correct reasoning traces for efficient data collection, and then designs a vision-aware attention score to quantify the model’s focus on visual information. Extensive experiments show that VISTA can be applied to various post-training scenarios, i.e., supervised fine-tuning and preference learning, and effectively enhances the multimodal reasoning performance across various MLLMs and tasks, e.g., bringing up to +13.66% average performance gains for Qwen2.5-VL-3B-Instruct.

The annealing phase is a pivotal convergence stage in LLM pre-training that ultimately determines final model quality. However, effectively selecting training data during this phase remains a key challenge. Current strategies rely on empirical heuristics, such as domain filtering or context extension, which lack a principled grounding in optimization theory. In this work, we characterize the annealing phase through the lens of the loss landscape's spectral geometry. We argue that optimal convergence requires gradient updates to satisfy heterogeneous constraints across different eigen-directions. Building on this insight, we formulate data selection as a problem of satisfying these directional constraints. To this end, we propose **DiReCT** (**Di**rectionally-**Re**strained **C**onstrained **T**raining), a novel framework that reformulates sample selection in the annealing stage as a constrained optimization problem. By imposing explicit directional constraints on per-sample gradients based on the spectral properties of the Hessian, **DiReCT** identifies samples that align with the optimal curvature-aware descent path. Extensive experiments across various model scales demonstrate that **DiReCT** consistently achieves state-of-the-art performance. For future research, code is available at https://github.com/xuyj233/Direct.

Evaluating the real-world capabilities of AI systems requires grounding benchmark performance in human-interpretable measures of task difficulty. Existing approaches that rely on direct human task completion time annotations are costly, noisy, and difficult to scale across benchmarks. In this work, we propose BRIDGE, a unified psychometric framework that learns the latent difficulty scale from model responses and anchors it to human task completion time. Using a two-parameter logistic Item Response Theory model, we jointly estimate latent task difficulty and model capability from model performance data across multiple benchmarks. We demonstrate that latent task difficulty varies linearly with the logarithm of human completion time, allowing human task completion time to be inferred for new benchmarks from model performance alone. Leveraging this alignment, we forecast frontier model capabilities in terms of human task length and independently reproduce METR’s exponential scaling results, with the 50% solvable task horizon doubling approximately every 6 months.