ICML 2026 Papers

Audio agents are increasingly deployed to execute tools from spoken requests, yet audio tool use poses challenges beyond text-only function calling: perception errors (e.g., homophones, noise, disfluencies) can corrupt entities and arguments, and natural interactions often require clarification that changes the tool-calling protocol. We introduce BFCL Audio, a large-scale benchmark for audio function calling with 6.2K expert-verified tasks across two suites that mirror common deployments: BFCL Text Audio (pipelined $\text{ASR}\rightarrow\text{LLM}\rightarrow\text{tools}$ via transcripts) and BFCL True Audio (end-to-end $\text{audio-in}\rightarrow\text{tool calls}$). BFCL Audio includes controlled speech and acoustic perturbations (accent and speaking-rate variation, content disfluencies, and background noise) generated through a controllable audio synthesis/augmentation pipeline. We provide automatic grading for both function names and argument values using AST-based matching for single-turn calls and response/state-based metrics for multi-turn interactions, enabling scalable evaluation without LLM judges. Across a broad set of models, we propose a failure-mode taxonomy and analyze which speech and noise factors most strongly impact tool-calling accuracy. We release the benchmark, evaluation harness, and audio pipeline to support research on reliable speech-based agents.

Visual AutoRegressive modeling (VAR) suffers from substantial computational cost due to the massive token count involved. Failing to account for the continuous variation in modeling, existing VAR token reduction methods face three key limitations: heuristic stage partition, non-adaptive schedules, and limited acceleration scope, thereby leaving significant acceleration potential untapped. Since entropy variation intrinsically reflects the transition of predictive uncertainty, it offers a principled measure to capture continuous modeling variation. Therefore, we propose NOVA, a training-free token reduction acceleration framework for VAR models via entropy analysis. NOVA adaptively determines the acceleration activation scale during inference by online identifying the inflection point of scale entropy growth. Through scale-linkage and layer-linkage ratio adjustment, NOVA dynamically computes distinct token reduction ratios for each scale and layer, pruning low-entropy tokens while reusing the cache derived from the residuals at the prior scale to accelerate inference and maintain generation quality. Extensive experiments and analyses validate NOVA as a simple yet effective training-free acceleration framework. Code is available.

The longitudinal management of blinding fundus diseases constitutes a Partially Observable Markov Decision Process (POMDP) necessitating a critical precision-risk trade-off between intervention and over-treatment, as true pathology is often obscured in static observations. However, existing paradigms fail to address this complexity. Traditional vision models remain uninterpretable and memoryless, and while Vision-Language Models (VLMs) excel in semantic understanding, they rely on unsafe open-loop text reasoning lacking the anatomical grounding essential for clinical safety. Furthermore, robust learning is hindered by the scarcity of process supervision in sparse clinical records. To bridge this gap, we introduce the Logic-Constrained Abductive Data Engine. Operating on a ``Propose-and-Verify'' paradigm, it validates MLLM-proposed biomarkers against clinical and temporal logic to reconstruct dense pathological states from sparse outcomes. Building on this foundation, we propose ORBIT, the first ophthalmic Prognostic World Model. Uniquely, ORBIT employs counterfactual visual foresight to imagine anatomical futures under different treatments, anchoring decisions in Closed-Loop Anatomical Verification rather than linguistic probabilities. Experiments demonstrate that ORBIT effectively captures disease evolution and establishes a new paradigm for human-in-the-loop longitudinal decision support and anatomically grounded treatment planning.

Large language models (LLMs) achieve remarkable performance through ever-increasing parameter counts, but scaling incurs steep computational costs. To better understand LLM scaling, we study representational differences between LLMs and their smaller counterparts, with the goal of replicating the representational qualities of larger models in smaller models. We observe a geometric phenomenon which we term $\textbf{embedding condensation}$, where token embeddings collapse into a narrow cone-like subspace in some language models. Through systematic analyses across multiple Transformer families, we show that small models such as $\texttt{GPT2}$ and $\texttt{Qwen3-0.6B}$ exhibit severe condensation, whereas larger models such as $\texttt{GPT2-xl}$ and $\texttt{Qwen3-32B}$ are more resistant to this phenomenon. Additional observations show that embedding condensation is not reliably mitigated by knowledge distillation from larger models. To fight against it, we formulate a dispersion loss that explicitly encourages embedding dispersion during training. Experiments demonstrate that it mitigates condensation, recovers dispersion patterns seen in larger models, and yields performance gains across 10 benchmarks. We believe this work offers a principled path toward improving smaller Transformers without additional parameters.

Proactive large language model (LLM) agents aim to actively plan, query, and interact over multiple turns, enabling efficient task completion beyond passive instruction following and making them essential for real-world, user-centric applications. Agentic reinforcement learning (RL) has recently emerged as a promising solution for training such agents in multi-turn settings, allowing interaction strategies to be learned from feedback. However, existing pipelines face a critical challenge in balancing task performance with user engagement, as passive agents can not efficiently adapt to users' intentions while overuse of human feedback reduces their satisfaction. To address this trade-off, we propose BAO, an agentic RL framework that combines behavior enhancement to enrich proactive reasoning and information-gathering capabilities with behavior regularization to suppress inefficient or redundant interactions and align agent behavior with user expectations. We evaluate BAO on multiple tasks from the UserRL benchmark suite, and demonstrate that it substantially outperforms RL baselines under controlled comparisons, while achieving comparable or even superior performance to frontier LLM agents, highlighting its effectiveness for training proactive, user-aligned LLM agents in complex multi-turn scenarios.

Foundation models have driven rapid progress in computer vision, yet the two dominant paradigms, vision-language foundation models (VLMs) and vision-only foundation models (VFMs), remain only partially compatible. VLMs offer language-grounded semantic alignment but are often visually coarse, while VFMs learn discriminative perceptual geometry but lack semantic grounding. We propose **GPUA**, a *Geometry-Preserving Unsupervised Alignment* framework that integrates the complementary strengths of VFMs and VLMs. Inspired by cross-lingual alignment, **GPUA** treats VFM features as a *visual language* and learns an *orthogonal* mapping that translates the VFM space into the VLM semantic space, preserving geometry and narrowing the modality gap without labels or model parameter updates. **GPUA** is task-agnostic and requires only feature-level access to pretrained models. Experiments across diverse benchmarks demonstrate improved cross-model compatibility and strong gains in downstream zero-shot recognition and segmentation with negligible overhead. Our code is available at: https://github.com/Yuteam14/GPUA.

Chain-of-thought (CoT) reasoning has become a central mechanism for eliciting multi-step reasoning in Large Language Models (LLMs). Yet recent evidence presents a tension: hidden states appear to already encode future reasoning before CoT fully unfolds, while explicit steps still remain crucial for tasks requiring compositional computation. To deepen the understanding between LLM's internal states and its verbalized reasoning trajectories, we investigate the latent planning strength of LLMs, through our probing method, Tele-Lens, applying to hidden states across diverse task domains. Our empirical results indicate that LLMs exhibit a myopic horizon, primarily conducting incremental transitions without precise global planning. Leveraging this characteristic, we propose a hypothesis on enhancing uncertainty estimation of CoT, which we validate that a sparse set of pivot positions can effectively represent the uncertainty of the entire path. We further underscore the significance of exploiting CoT dynamics, and demonstrate that automatic recognition of CoT bypass can be achieved without performance degradation. Our code, data and models are released at https://github.com/lxucs/tele-lens.

Large-scale, multi-subject widefield calcium imaging provides unprecedented access to brain-wide cortical dynamics. However, the high dimensionality, complex spatiotemporal structure, and substantial task-irrelevant activity in widefield recordings have largely restricted modeling efforts to single-session analyses, limiting scalability and generalization. While multi-subject pretrained models have been explored for some neural modalities, multi-subject models for widefield calcium imaging have not yet been demonstrated; further, subject-invariant zero-shot behavior decoding remains elusive for multi-subject models across neural modalities more broadly. As a first step toward foundation modeling of widefield data, we introduce WiCAT, a multi-subject model that leverages self-supervised pretraining to both outperform single-session models and enable zero-shot behavior decoding on unseen subjects. WiCAT introduces an atlas-grounded tokenization scheme without session-specific components and learns globally shared spatiotemporal representations. Across multiple widefield datasets, the pretrained model supports lightweight downstream decoding, transfers across subjects, tasks, and datasets, and outperforms baseline models. Notably, the model also achieves robust zero-shot continuous behavior decoding and left-out brain region reconstruction on unseen subjects.

Leveraging the unlabeled stream is crucial yet challenging in Semi-Supervised Continual Learning (SSCL) under continual class expansion. Existing SSCL methods typically enforce dense pseudo-label consistency and indiscriminate distillation on unlabeled data, which can reinforce errors and intensify base–novel interference. To address these issues, we propose Discrete-anchored Incremental Learning (DiL) to ground continual updates on reliable discrete anchors that remain stable under noisy pseudo-labels. DiL introduces Discrete Contrastive Distillation (DCD), which discretizes the distillation pathway and performs anchor-referenced selective distillation to curb error reinforcement. Meanwhile, Class-Aware Channel-Chunked Encoding (CACE) learns channel-chunked representations and exploits the confusion matrix induced by the discrete anchors to separate novel from confusable base classes. Extensive experiments on multiple datasets show that DiL achieves state-of-the-art performance across diverse SSCL protocols.

Decentralized SGD is a fundamental algorithm in decentralized learning, although the influence of an underlying network topology on its convergence behavior is not yet fully understood. Existing convergence analyses have shown that topologies with a small spectral gap significantly deteriorate the convergence rate of Decentralized SGD in both homogeneous and heterogeneous cases. However, many prior papers have reported that indeed the choice of the topology has a significant experimental impact in the heterogeneous case, but has little experimental impact on training behavior in the homogeneous case. In this paper, we present a tighter convergence analysis of Decentralized SGD, offering a more precise understanding of how topologies affect the convergence rate than the prior analysis. Specifically, unlike existing convergence analyses that used only the spectral gap as a property of the topology, our novel analysis shows that all eigenvalues of the mixing matrix affect the convergence rate. Throughout the experiments, we carefully evaluated the convergence behavior of Decentralized SGD and demonstrated that our novel convergence analysis can more accurately describe the effect of topology on the convergence rate.

3D editing—the task of locally modifying the geometry or appearance of a 3D asset—has wide applications in immersive content creation, digital entertainment, and AR/VR. However, unlike 2D editing, it remains challenging due to the need for cross-view consistency, structural fidelity, and fine-grained controllability. Existing approaches are often slow, prone to geometric distortions, or dependent on manual and accurate 3D masks that are error-prone and impractical. To address these challenges, we advance both the data and model fronts. On the data side, we introduce 3DEditVerse, the largest paired 3D editing benchmark to date, comprising 116,309 high-quality training pairs and 1,500 curated test pairs. Built through complementary pipelines of pose-driven geometric edits and foundation model-guided appearance edits, 3DEditVerse ensures edit locality, multi-view consistency, and semantic alignment. On the model side, we propose 3DEditFormer, a 3D-structure-preserving transformer. By enhancing image-to-3D generation with dual-guidance attention and time-adaptive gating, 3DEditFormer disentangles editable regions from preserved structure, enabling precise and consistent edits without requiring auxiliary 3D masks. Extensive experiments demonstrate that our framework outperforms state-of-the-art baselines both quantitatively and qualitatively, establishing a new standard for practical and scalable 3D editing. Dataset and code are available at https://www.lv-lab.org/3DEditFormer/

Large language model (LLM)-based multi-agent systems commonly rely on natural-language prompts to specify agent behavior, yet optimizing these prompts remains challenging when agent roles and interaction structures are fixed by design. In such systems, behaviors emerge over long, noisy interaction trajectories, making it difficult to determine which prompt components are responsible for success or failure. As a result, outcome-level feedback alone is insufficient, while existing prompt optimization methods typically rely on final task scores or global prompt rewrites, limiting their ability to exploit trajectory evidence or support the localized updates. We propose Trajectory-based Rule Credit Estimation (TRUCE), a framework for prompt optimization in multi-agent systems that explicitly addresses this credit assignment challenge. TRUCE performs trajectory-aware attribution by linking outcome feedback to informative sub-trajectories and translating the resulting credit signals into unit-level edits over prompt-defined behavioral rules. By preserving agent roles and interaction structures, TRUCE enables prompt refinement through localized updates aggregated across tasks. Experiments on multiple benchmarks demonstrate that TRUCE consistently improves task performance and efficiency over competitive baselines. Code is available at https://github.com/bingo-w/TRUCE.

Diffusion language models (dLLMs) recently emerged as a promising alternative to auto-regressive LLMs. The latest works further extended it to multimodal understanding and generation tasks. In this work, we propose LaViDa-R1, a multimodal, general-purpose reasoning dLLM. Unlike existing works that build reasoning dLLMs through task-specific reinforcement learning, LaViDa-R1 incorporates diverse multimodal understanding and generation tasks in a unified manner. In particular, LaViDa-R1 is built with a novel unified post-training framework that seamlessly integrates supervised finetuning (SFT) and multi-task reinforcement learning (RL). It employs several novel training techniques, including answer-forcing, tree search, and complementary likelihood estimation, to enhance effectiveness and scalability. Extensive experiments demonstrate LaViDa-R1's strong performance on a wide range of multimodal tasks, including visual math reasoning, reason-intensive grounding, and image editing.

In scene graph generation, a central challenge is modeling polysemous predicates whose meanings shift across contexts. Prior approaches address this issue by decomposing predicates into multiple static prototypes or retrieving semantically similar exemplars. However, these strategies keep predicate representations static and cannot reorganize semantics to reflect image-specific evidence, leading to systematic confusions in ambiguous contexts. We propose **AlignG**, which learns context-conditioned predicate semantics via prototype feedback. AlignG infers context-conditioned predicate semantics from the relation candidates within each image and feeds the adapted semantics back to recalibrate relation representations. The learning objective anchors this adaptation to global semantic centers, preventing semantic drift while still allowing selective reorganization when the scene provides consistent relational cues. Experiments on VG-150 and GQA-200 show consistent improvements over state-of-the-art baselines, with F@100 improvements of +1.4 on VG-150 and +2.7 on GQA-200 under SGDet. We further visualize per-image prototype similarity shifts and observe coherent context-dependent reorganization where prototypes selectively merge or separate predicates according to scene evidence. The code is available at https://github.com/Namgyu97/AlignG-SGG.pytorch.

The growing capabilities of large language models (LLMs) have driven their widespread deployment across diverse domains, even in potentially high-risk scenarios. Despite advances in safety alignment techniques, current models remain vulnerable to emerging *persona-based jailbreak attacks*. Existing research on persona-based jailbreak has primarily focused on attack iterations, yet it lacks systemic and mechanistic constraints on the defense side. To address this challenge, we propose Persona-Invariant Alignment (PIA), an adversarial self-play framework that achieves co-evolution through Persona Lineage Evolution (PLE) on the attack side and Persona-Invariant Consistency Learning (PICL) on the defense side. Theoretically, PICL is grounded in the *structural separation hypothesis*, using a unilateral KL-divergence constraint to enable the structural decoupling of safety decisions from persona context, thereby maintaining safe behavior under persona-based jailbreak attacks. Experimental results demonstrate that PLE efficiently explores high-risk persona spaces by leveraging lineage-based credit propagation. Meanwhile, the PICL defense method significantly reduces the Attack Success Rate (ASR) while preserving the model's general capability, thereby validating the superiority and robustness of this alignment paradigm. WARNING: This paper contains potentially offensive and harmful text.

Decentralized class continual learning refers to a paradigm where distributed clients continuously acquire new classes while retaining previously learned information without relying on a central server. With increasing emphasis on privacy preservation, there is a growing need for on-demand unlearning, introducing two key challenges: Historical Class Unlearning and Network-Wide Knowledge Entanglement. In this work, we propose a decentralized continual learning framework with on-demand unlearning (DCU), which is the first attempt at achieving class continual learning and arbitrary-time class unlearning in a distributed setting. Specifically, our proposed DCU comprises three main stages: prototypes extraction, prototype-guided continual learning, and unlearning with disposable prototypes. Firstly, the prototypes extraction mechanism is designed to capture the class-specific concepts as lightweight, disposable embeddings. Then, the synthetic data guided by these prototypes can be combined with real data to achieve incremental learning through distillation. Besides, synthetic samples with noisy label are used to guide the adjustment of the model's decision boundary, effectively erasing the influence of the target class while preserving other classes' knowledge. Extensive experiments conducted on two datasets demonstrate the effectiveness of our DCU in dynamic learning and target class unlearning.

Training Large Language Models (LLMs) to reason often relies on Reinforcement Learning (RL) with task-specific verifiers. However, many real-world reasoning-intensive tasks lack verifiers, despite offering abundant expert demonstrations that remain under-utilized for reasoning-focused training. We introduce **RARO** (Relativistic Adversarial Reasoning Optimization), which learns strong reasoning capabilities from expert demonstrations alone via **Inverse Reinforcement Learning**. Our method sets up an adversarial game between a **policy** and a **relativistic critic**: the policy learns to mimic expert answers, while the critic aims to identify the expert among (expert, policy) answer pairs. Both the policy and the critic are trained jointly and continuously via RL, and we identify key stabilization techniques required for robust learning. Empirically, RARO significantly outperforms strong verifier-free baselines across all evaluation tasks: $+13.7\%$ accuracy on Countdown ($1.5$B), +8.2\% on DeepMath ($7$B), and $+19.1\%$ win-rate on Poetry Writing ($7$B) against expert poems. RARO also exhibits similar robust scaling trends as RL with verifiers. These results demonstrate that RARO effectively elicits strong reasoning performance from expert demonstrations alone, enabling robust reasoning learning even when task-specific verifiers are unavailable.

Recent years have witnessed significant progress in developing effective diffusion models. Parallel sampling is a promising recent approach that reformulates the sequential denoising process as solving a system of nonlinear equations, and it can be combined with other acceleration techniques. However, current progress is limited by the trade-off between high fidelity and computational efficiency. This paper addresses the challenge of scaling to high-dimensional, multi-modal generation. Specifically, we present ROPA (Robust Parallel Diffusion Sampling), which takes into account the properties of the denoising process and solves the linear system using adaptive local sparsity to achieve stable parallel sampling. Extensive experiments demonstrate ROPA’s effectiveness: it significantly accelerates sampling across diverse image and video diffusion models, achieving up to $2.9\times$ speedup with eight core, an improvement of 20.8\% over baselines without sacrificing sample quality. ROPA enables parallel sampling methods to provide a solid foundation for real-time, high-fidelity diffusion generation.

We investigate the role of network architecture in shaping the inductive biases of modern score-based generative models. To this end, we introduce the Score Anisotropy Directions (SADs), architecture-dependent directions that reveal how different networks preferentially capture data structure. Our analysis suggests that SADs form adaptive bases aligned with the architecture's output geometry, providing a principled way to predict generalization ability in score models prior to training. Through both synthetic data and standard image benchmarks, we demonstrate that SADs reliably capture fine-grained model behavior and correlate with downstream performance, as measured by Wasserstein metrics. Our work offers a new lens for explaining and predicting directional biases of generative models.