ICLR 2026 论文列表

The success of LLMs has sparked interest in various agentic applications. A key hypothesis is that LLMs, leveraging common sense and Chain-of-Thought (CoT) reasoning, can effectively explore and efficiently solve complex domains. However, LLM agents have been found to suffer from sub-optimal exploration and the knowing-doing gap, the inability to effectively act on knowledge present in the model. In this work, we systematically study why LLMs perform sub-optimally in decision-making scenarios. In particular, we closely examine three prevalent failure modes: greediness, frequency bias, and the knowing-doing gap. We propose mitigation of these shortcomings by fine-tuning via Reinforcement Learning (RL) on self-generated CoT rationales. Our experiments across multi-armed bandits, contextual bandits, and Tic-tac-toe demonstrate that RL fine-tuning enhances the decision-making abilities of LLMs by increasing exploration and narrowing the knowing-doing gap. Finally, we study both classic exploration mechanisms, such as $\epsilon$-greedy, and LLM-specific approaches, such as self-correction and self-consistency, to enable more effective fine-tuning of LLMs for decision-making.

We present DistillKac, a fast image generator that uses the damped wave equation and its stochastic Kac representation to move probability mass at finite speed. In contrast to diffusion models whose reverse time velocities can become stiff and implicitly allow unbounded propagation speed, Kac dynamics enforce finite speed transport and yield globally bounded kinetic energy. Building on this structure, we introduce classifier free guidance in velocity space that preserves square integrability under mild conditions. We then propose endpoint only distillation that trains a student to match a frozen teacher over long intervals. We prove a stability result that promotes supervision at the endpoints to closeness along the entire path. Experiments demonstrate DistillKac delivers high quality samples with very few function evaluations while retaining the numerical stability benefits of finite speed probability flows.

Large Language Models (LLMs) are increasingly employed in real-world applications, driving the need to evaluate the trustworthiness of their generated text. To this end, reliable uncertainty estimation is essential. Leading uncertainty estimation methods generate and analyze multiple output sequences, which is computationally expensive and impractical at scale. In this work, we inspect the theoretical foundations of these methods and explore new directions to enhance computational efficiency. Building on the framework of proper scoring rules, we find that the negative log-likelihood of the most likely output sequence constitutes a theoretically principled uncertainty measure. To approximate this alternative measure, we propose G-NLL, obtained using a single output sequence from greedy decoding. This approach streamlines uncertainty estimation while preserving theoretical rigor. Empirical results demonstrate that G-NLL achieves state-of-the-art performance across various scenarios. Our work lays the theoretical foundation for efficient and reliable uncertainty estimation in natural language generation, challenging the necessity of the prevalent methods that are more complex and resource-intensive.

Learning in the combinatorially large output space of sequence generation problems is challenging as providing expert demonstrations scales poorly with sequence length, and RL struggles with sparse rewards. Between dense demonstrations in supervised training and no demonstrations in reinforcement learning lies an underexplored regime: partial supervision. We ask whether some classes of sequence learning problems become efficiently learnable by exploiting this gap. We address this by introducing adaptive backtracking (AdaBack), a per-sample curriculum learning algorithm that reveals a partial prefix of the target output. The supervision length is adjusted dynamically for each sample based on the model’s past reward signal, allowing it to incrementally learn to complete reasoning chains by conditioning on correct partial solutions. We investigate this intermediate regime between SFT and RL and argue that per-sample curriculum learning is more than a trade-off between efficiency and generality—it can succeed in tasks with long sequences of latent dependencies where SFT and RL both fail to generalize. Using a synthetic task with latent parity constraints, we show that AdaBack reliably solves problems that are otherwise intractable. On two mathematical reasoning benchmarks, MATH and GSM8k, we find that AdaBack enables models to solve problems that RL alone cannot, acquiring new reasoning capabilities through incremental exposure to partial solutions.

Egocentric vision is essential for both human and machine visual understanding, particularly in capturing the detailed hand-object interactions needed for manipulation tasks. Translating third-person views into first-person views significantly benefits augmented reality (AR), virtual reality (VR) and robotics applications. However, current exocentric-to-egocentric translation methods are limited by their dependence on 2D cues, synchronized multi-view settings, and unrealistic assumptions such as the necessity of an initial egocentric frame and relative camera poses during inference. To overcome these challenges, we introduce *EgoWorld*, a novel two-stage framework that reconstructs an egocentric view from rich exocentric observations, including projected point clouds, 3D hand poses, and textual descriptions. Our approach reconstructs a point cloud from estimated exocentric depth maps, reprojects it into the egocentric perspective, and then applies diffusion-based inpainting to produce dense, semantically coherent egocentric images. Evaluated on 4 datasets (i.e., H2O, TACO, Assembly101, and Ego-Exo4D), *EgoWorld* achieves state-of-the-art performance and demonstrates robust generalization to new objects, actions, scenes, and subjects. Moreover, *EgoWorld* exhibits robustness on in-the-wild examples, underscoring its practical applicability.

Emotional information in speech plays a unique role in multimodal perception. However, current Speech Large Language Models (SpeechLLMs), similar to conventional speech emotion recognition (SER) systems, still treat emotion understanding as a simple classification problem. This provides limited interpretability of predictions, while leaving the LLMs’ expressive and reasoning capabilities underutilized. In this work, we take the first step to reformulate SER as a deep reasoning problem through reinforcement learning (RL). We propose EmotionThinker, which is designed to generate accurate emotion predictions with interpretable explanations grounded in fine-grained acoustic cues. To achieve this, we first construct EmotionCoT-35K, an emotional reasoning dataset with Chain-of-Thought annotations and detailed captions. Second, we observe that current SpeechLLMs exhibit weak prosody perception, whereas prosodic cues constitute fundamental signals for interpreting emotions. To address this, we develop the prosody-enhanced foundation model EmotionThinker-Base, and demonstrate that prosody enhancement improves emotion understanding. Third, we introduce Group-Relative-Policy-Optimization with Progressive-Trust-aware-Reasoning-Reward (GRPO-PTR}) for RL. Different from standard GRPO, which relies only on rule-based outcome rewards, GRPO-PTR progressively introduces reasoning reward, dynamically adjusts it with a trustworthiness weight reflecting the alignment between reasoning and outcome, and evaluates the overall reasoning quality with a reward model based on multi-dimensional criteria. EmotionThinker outperforms previous state-of-the-art evaluation models both in emotion accuracy and explanation quality, advancing SER toward interpretable multimodal reasoning.

Methods addressing Learning with Noisy Labels (LNL) and multi-source Domain Generalization (DG) use training techniques to improve downstream task performance in the presence of label noise or domain shifts, respectively. Prior work often explores these tasks in isolation, with only limited work that evaluates how label noise affects existing DG methods without also exploring methods to reduce its effect. However, many applications require methods that are robust to both label noise and distribution shifts, which we refer to as Noise-Aware Generalization (NAG), and when these problems are considered together new challenges emerge. E.g., most LNL methods identify noise by detecting distribution shifts in a class’s samples, i.e., they assume that distribution shifts often correspond to label noise. In NAG distribution shifts can be due to label noise or domain shifts, breaking the assumptions used by LNL methods. DG methods often overlook the effect of label noise entirely, which can confuse a model during training, reducing performance. A naive solution to this issue is to make a similar assumption as many DG methods, where we presume to have domain labels during training, enabling us to isolate the two types of shifts. However, this ignores valuable cross-domain information. Specifically, our proposed DL4ND approach improves noise detection by taking advantage of the observation that noisy samples that may appear indistinguishable within a single domain often show greater variation when compared across domains. Experiments show DL4ND significantly improves performance across seven diverse datasets, offering a promising direction for tackling NAG.

While Large Language Models (LLMs) show immense promise as planners for embodied AI, their stochastic nature and lack of formal reasoning capabilities prevent the strict safety guarantees required for physical deployment. Current approaches fall short: they either rely on other unreliable LLMs for safety checks or simply reject unsafe plans without offering a path to success. This work bridges this critical gap by introducing the Verifiable Iterative Refinement Framework (VIRF), a neuro-symbolic architecture that shifts the paradigm from a passive safety gatekeeper to an active safety collaborator. Where prior verifiers simply reject failures, our framework provides causal, pedagogical feedback that teaches the LLM why its plan was unsafe, enabling intelligent repairs rather than mere avoidance.Our core contribution is a novel tutor-apprentice dialogue, where a deterministic Logic Tutor, grounded in a formal safety ontology, provides causal and explanatory feedback to an LLM Apprentice planner. This pedagogical interaction allows the apprentice to perform intelligent, creative plan repairs, resolving safety conflicts rather than merely avoiding them. To ground this dialogue in verifiable truth, we introduce a scalable knowledge acquisition pipeline that synthesizes a comprehensive safety knowledge base from real-world documents, a process that simultaneously reveals and corrects significant blind spots in existing benchmarks. On a new suite of challenging home safety tasks, VIRF achieves a perfect 0\% Hazardous Action Rate (HAR), completely eliminating unsafe actions while attaining a 77.3\% Goal-Condition Rate (GCR)—the highest among all baselines. It does so with remarkable efficiency, requiring only 1.1 correction iterations on average. By acting as a verifiable safety scaffold, VIRF demonstrates a principled and robust pathway toward building embodied agents that are not just capable, but fundamentally trustworthy.

A central challenge in large language model inference is the trade-off between generation speed and output quality. Autoregressive models produce high-quality text but generate tokens sequentially. Diffusion models can generate tokens in parallel but often need many iterations to match the same quality. We propose planned diffusion, a hybrid method that combines the strengths of both paradigms. Planned diffusion works in two stages: first, the model creates a short autoregressive plan that breaks the output into smaller, independent spans. Second, the model generates these spans simultaneously using diffusion. This approach expands the speed–quality Pareto frontier and provides a practical path to faster, high-quality text generation. On AlpacaEval, a suite of 805 instruction-following prompts, planned diffusion achieves Pareto-optimal trade-off between quality and latency, achieving 1.27x to 1.81x speedup over autoregressive generation with only 0.87\% to 5.4\% drop in win rate, respectively. Our sensitivity analysis shows that the planning mechanism of planned diffusion is minimal and reliable, and simple runtime knobs exist to provide flexible control of the quality-latency trade-off.

Data quality plays a critical role in enhancing supervised fine-tuning (SFT) for large language models (LLMs), and token-level data selection has emerged as a promising direction for its fine-grained nature. Despite their strong empirical performance, existing token-level selection methods share two key limitations: (1) requiring training or accessing an additional reference model, and (2) relying solely on loss information for token selection, which cannot well preserve semantically important tokens that are not favored by loss-based metrics. To address these challenges, we propose **ssToken**, a **S**elf-modulated and **S**emantic-aware **Token** Selection approach. ssToken leverages readily accessible history models to compute the per-token loss difference with the current model, which serves as a self-modulated signal that enables the model to adaptively select tokens along its optimization trajectory, rather than relying on excess loss from an offline-trained reference model as in prior works. We further introduce a semantic-aware, attention-based token importance estimation metric, orthogonal to loss-based selection and providing complementary semantic information for more effective filtering. Extensive experiments across different model families and scales demonstrate that both self-modulated selection and semantic-aware selection alone outperform full-data fine-tuning, while their integration—ssToken—achieves synergistic gains and further surpasses prior token-level selection methods, delivering performance improvements while maintaining training efficiency. Source code is available at https://anonymous.4open.science/r/Submission2116-B7C5.

Large language models (LLMs) have been widely deployed with rapidly expanding context windows to support increasingly demanding applications. However, long contexts pose significant deployment challenges, primarily due to the KV cache whose size grows proportionally with context length. While KV cache compression methods are proposed to address this issue, KV dropping methods incur considerable accuracy loss, and KV retrieval methods suffer from significant efficiency bottlenecks. We propose FreeKV, an algorithm-system co-optimization framework to enhance KV retrieval efficiency while preserving accuracy. On the algorithm side, FreeKV introduces speculative retrieval to shift the KV selection and recall processes out of the critical path, combined with fine-grained correction to ensure accuracy. On the system side, FreeKV employs hybrid KV layouts across CPU and GPU memory to eliminate fragmented data transfers, and leverages double-buffered streamed recall to further improve efficiency, enabling effective overlap with computation, full latency hiding, and practical speedups from speculative recall. Experiments demonstrate that FreeKV achieves near-lossless accuracy across various scenarios and models, delivering up to 13$\times$ speedup compared to SOTA KV retrieval methods.

Multimodal models have achieved remarkable progress, yet recent studies suggest they struggle with compositional reasoning, often performing at or below random chance on established benchmarks. We revisit this problem and show that widely used evaluation metrics systematically underestimate model capabilities. To address this, we introduce a group matching score that better leverages group structure and uncovers substantial hidden competence in both contrastive vision–language models (VLMs) and multimodal large language models (MLLMs). Moreover, simply overfitting to the induced group matchings at test time transfers this hidden competence into higher scores under the original evaluation metric, closing much of the reported gap. With this adjustment, GPT-4.1 becomes the first system to surpass estimated human performance on Winoground. Building on this insight, we propose Test-Time Matching (TTM), an iterative self-training algorithm that bootstraps model performance without any external supervision. TTM delivers further non-trivial improvements: for example, SigLIP-B16 with TTM surpasses GPT-4.1 on MMVP-VLM, establishing a new state of the art. Importantly, TTM is broadly effective even on benchmarks without metric-induced effects or group structures, achieving relative gains exceeding 85.7\% on challenging datasets such as Whatsup. Across 16 datasets and variants, our experiments consistently demonstrate that TTM unlocks hidden compositional reasoning ability and advances the frontier of multimodal evaluation.

Flow-matching models deliver state-of-the-art fidelity in image and video generation, but the inherent sequential denoising process renders them slower. Existing acceleration methods like distillation, trajectory truncation, and consistency approaches are static, require retraining, and often fail to generalize across tasks. We propose FastFlow, a plug-and-play adaptive inference framework that accelerates generation in flow matching models. FastFlow identifies denoising steps that produce only minor adjustments to the denoising path and approximates them without using the full neural network models used for velocity predictions. The approximation utilizes finite-difference velocity estimates from prior predictions to efficiently extrapolate future states, enabling faster advancements along the denoising path at zero compute cost. This enables skipping computation at intermediary steps. We model the decision of how many steps to safely skip before requiring a full model computation as a multi-armed bandit problem. The bandit learns the optimal skips to balance speed with performance. FastFlow integrates seamlessly with existing pipelines and generalizes across image generation, video generation, and editing tasks. Experiments demonstrate a speedup of over $2.6\times$ while maintaining high-quality outputs. The source code for this work can be found at https://github.com/Div290/FastFlow.

Autoregressive large language models (LLMs) have unified a vast range of language tasks, inspiring preliminary efforts in autoregressive (AR) video generation. Existing AR video generators either diverge from standard LLM architectures, depend on bulky external text encoders, or incur prohibitive latency due to next-token decoding. In this paper, we introduce Lumos-1, an LLM-based unified model for AR video generation with efficient discrete diffusion. Firstly, to fit videos with LLMs, we identify that 1D RoPE is ill-suited for visual spatiotemporal correlation modeling, and while demonstrated to be useful, naive 3D RoPE exhibits imbalanced frequency spectra. Therefore, we propose MM‑RoPE, which preserves the original textual RoPE while seamlessly accommodating video data with comprehensive frequency spectra and scaled 3D positions. Secondly, to fit the video data's nature and overcome the inefficiency of next-token decoding, we adopt a parallel and mask-based discrete diffusion with the intra-frame bidirectional and inter-frame causal attention masks. Based on this attention mask, we uncover the frame‑wise loss imbalance issue caused by spatial information redundancy and propose Autoregressive Discrete Diffusion Forcing, which introduces temporal tube masking during training with a compatible inference‑time masking policy to avoid quality degradation. Despite using only 48 GPUs for pre-training, limited data and a discrete tokenizer, Lumos-1 achieves results surpassing those of Show-o2 on GenEval, COSMOS-Video2World on VBench-I2V, and OpenSoraPlan on VBench-T2V.

Next token prediction has been the standard training objective used in large language model pretraining. Representations are learned as a result of optimizing for token-level perplexity. We propose Continuous Concept Mixing (CoCoMix), a novel pretraining framework that combines discrete next token prediction with continuous concepts. Specifically, CoCoMix predicts ``continuous concepts'' learned from a pretrained sparse autoencoder and mixes them into the model's hidden state by interleaving with token hidden representations. Through experiments on multiple benchmarks, including language modeling and downstream reasoning tasks, we show that CoCoMix is more sample efficient and consistently outperforms standard next token prediction and knowledge distillation. We find that combining both concept learning and interleaving in an end-to-end framework is critical to performance gains. Furthermore, CoCoMix enhances interpretability and steerability by allowing direct inspection and modification of the predicted concept, offering a transparent way to guide the model’s internal reasoning process.

AI control protocols serve as a defense mechanism to stop untrusted LLM agents from causing harm in autonomous settings. Prior work treats this as a security problem, stress testing with exploits that use the deployment context to subtly complete harmful side tasks, such as backdoor insertion. In practice, most AI control protocols are fundamentally based on LLM monitors, which can become a central point of failure. We study \textit{adaptive} attacks by an untrusted model that knows the protocol and the monitor model, which is plausible if the untrusted model was trained with a later knowledge cutoff or can search for this information autonomously. We instantiate a simple adaptive attack vector by which the attacker embeds known or zero-shot prompt injections in the model outputs. Using this tactic, frontier models consistently evade diverse monitors and complete malicious tasks on two main AI control benchmarks. The attack works universally against current protocols that rely on a monitor. Furthermore, the recent Defer-to-Resample protocol even backfires, as its resampling amplifies the prompt injection and effectively reframes it as a best-of-$n$ attack. In general, adaptive attacks on monitor models represent a major blind spot in current control protocols and should become a standard component of evaluations for future AI control mechanisms.

While diffusion models have achieved remarkable performance in image generation, they often struggle with the imbalanced datasets frequently encountered in real-world applications, resulting in significant performance degradation on minority classes. In this paper, we identify model capacity allocation as a key and previously underexplored factor contributing to this issue, providing a perspective that is orthogonal to existing research. Our empirical experiments and theoretical analysis reveal that majority classes monopolize an unnecessarily large portion of the model's capacity, thereby restricting the representation of minority classes. To address this, we propose Capacity Manipulation (CM), which explicitly reserves model capacity for minority classes. Our approach leverages a low-rank decomposition of model parameters and introduces a capacity manipulation loss to allocate appropriate capacity for capturing minority knowledge, thus enhancing minority class representation. Extensive experiments demonstrate that CM consistently and significantly improves the robustness of diffusion models on imbalanced datasets, and when combined with existing methods, further boosts overall performance.

As vision-language models (VLMs) gain prominence, their multimodal interfaces also introduce new safety vulnerabilities, making the safety evaluation challenging and critical. Existing red-teaming efforts are either restricted to a narrow set of adversarial patterns or depend heavily on manual engineering, lacking scalable exploration of emerging real-world adversarial strategies. To bridge this gap, we propose ARMs, an adaptive red-teaming agent that systematically conducts comprehensive risk assessments for VLMs. Given a target harmful behavior or risk definition, ARMs automatically optimizes diverse red-teaming strategies with reasoning-enhanced multi-step orchestration, to effectively elicit harmful outputs from target VLMs. This is the first red teaming framework that provides controllable generation given risk definitions. We propose 11 novel multimodal attack strategies, covering diverse adversarial patterns of VLMs (e.g., reasoning hijacking, contextual cloaking), and integrate 17 red-teaming algorithms with ARMs. To balance the diversity and effectiveness of the attack, we design a layered memory with an epsilon-greedy attack algorithm. Extensive experiments on different instance-based benchmarks and policy-based safety evaluations show that ARMs achieves the state-of-the-art attack success rate (ASR), improving ASR by an average of 52.1% compared to existing baselines and even exceeding 90% ASR on Claude-4-Sonnet, a constitutionally-aligned model widely recognized for its robustness. We show that the diversity of red-teaming instances generated by ARMs is significantly higher, revealing emerging vulnerabilities in VLMs. Leveraging ARMs, we construct ARMs-Bench, a large-scale multimodal safety benchmark comprising 30K red-teaming instances spanning 51 diverse risk categories, grounded in both real-world multimodal threats and regulatory risks. Fine-tuning with ARMs-Bench substantially reduces ASR while preserving general utility of VLMs, providing actionable insights to improve multimodal safety alignment.

We present **Lean4PHYS**, a comprehensive reasoning framework for college-level physics problems in Lean4. **Lean4PHYS** includes *LeanPhysBench*, a college-level benchmark for Lean4 formal physics reasoning, which contains 200 hand-crafted and peer-reviewed statements formalized from university textbooks and physics competition problems. To establish a solid foundation for formal reasoning in physics, we also launch *PhysLib*, a repository that contains fundamental unit systems and theorems essential for formal physics proving, which aims to be community-driven and long-term maintained. Based on the *LeanPhysBench* and *PhysLib* we composed in **Lean4PHYS**, we perform exhaustive experiments of baseline results using major expert Math Lean4 provers and state-of-the-art closed-source models, and provide an analysis of their performance. In the experiment, we identify that most expert provers do not outperform general models as they did in the math domain. This indicates a potential overfitting of the math domain rather than learning formal reasoning for formal provers. We also conduct a comprehensive experiment showing that with *PhysLib* in the context, LLMs' performance on *LeanPhysBench* increases by 11.90\% on average, proving the effectiveness of our repository in assisting LLMs to solve the Lean4 physics problem. To the best of our knowledge, we are the first study to provide a physics benchmark in Lean4.