ICML 2026 论文列表

Multimodal large language models (MLLMs) have demonstrated remarkable reasoning capabilities over internalized knowledge. However, current research overlooks contextual reasoning, the ability to reason based on the relevant information present in the context. To investigate this issue, we construct the Visual Contextual Reasoning Benchmark (ContextReasonV-Bench), and our analysis reveals two predominant failure modes: *context neglect*, where models rely on pre-trained knowledge instead of contextual information, and *superficial pattern matching*, where models exploit shallow correlations rather than genuine patterns. To address these limitations, we propose a two-stage approach that progressively establishes and reinforces contextual pattern acquisition. The first stage establishes an "analyze-then-solve" reasoning paradigm through supervised fine-tuning (SFT). We then employ a context-aware reinforcement learning (RL) framework that integrates context-aware reward modeling with hierarchical advantage estimation to encourage the model to identify genuine contextual patterns. This approach yields Context-Aware Reasoner (CAR), a model that achieves 38.14\% accuracy on ContextReasonV-Bench, a 22.09\% improvement over the base model, and generalizes effectively to unseen tasks.

Time-series anomaly detection (TSAD) with multimodal large language models (MLLMs) is an emerging area, yet a persistent challenge remains: MLLMs rely on coarse time-series heuristics but struggle with multi-dimensional, detailed reasoning, which is vital for understanding complex time-series data. We present AnomSeer to address this by reinforcing the model to ground its reasoning in precise, structural details of time series, unifying anomaly classification, localization, and explanation. At its core, an expert chain-of-thought trace is generated to provide verifiable, fine-grained reasoning from classical analyses (e.g., statistical measures, frequency transforms). Building on this, we propose a novel time-series grounded policy optimization (TimerPO) that incorporates two additional components beyond standard reinforcement learning: a time-series grounded advantage based on optimal transport and an orthogonal projection to ensure this auxiliary granular signal does not interfere with the primary detection objective. Across diverse anomaly scenarios, AnomSeer, with Qwen2.5-VL-3B/7B-Instruct, outperforms larger commercial baselines in classification and localization accuracy, particularly on point- and frequency-driven exceptions. Moreover, it produces plausible reasoning traces that support its conclusions.

Controllable diffusion generation often relies on various heuristics that are seemingly disconnected without a unified understanding. We bridge this gap with Diffusion Controller (DiffCon), a unified control-theoretic view that casts reverse diffusion sampling as state-only stochastic control within (generalized) linearly-solvable Markov Decision Processes (LS-MDPs). Under this framework, control acts by reweighting the pretrained reverse-time transition kernels, balancing terminal objectives against an $f$-divergence cost. From the resulting optimality conditions, we derive practical reinforcement learning methods for diffusion fine-tuning: (i) $f$-divergence-regularized policy-gradient updates, including a PPO-style rule, and (ii) a regularizer-determined reward-weighted regression objective with a minimizer-preservation guarantee under the Kullback–Leibler (KL) divergence. The LS-MDP framework further implies a principled model form: the optimal score decomposes into a fixed pretrained baseline plus a lightweight control correction, motivating a side-network parameterization conditioned on exposed intermediate denoising outputs, enabling effective *gray-box adaptation* with a frozen backbone. Experiments on Stable Diffusion v1.4 across supervised and reward-driven finetuning show consistent gains in preference-alignment win rates and improved quality–efficiency trade-offs versus gray-box baselines and even the parameter-efficient white-box adapter LoRA.

We address the challenge of automatically generating diverse attack prompts for large language models (LLMs) that elicit harmful behaviors (e.g., insults, sexual content) and are used for safety fine-tuning. While several prior approaches train LLMs with reinforcement learning (RL) to generate such prompts using only a toxicity classifier as a reward, existing diversity-seeking RL methods often collapse to limited modes: once high-reward prompts are found, exploration of new regions is discouraged. Inspired by the active learning paradigm that encourages adaptive exploration, we introduce \textbf{Active Attacks}, a novel RL-based red-teaming algorithm that adapts its attacks as the victim evolves. By periodically safety fine-tuning the victim LLM with collected attack prompts, we naturally induce an \emph{easy-to-hard exploration curriculum}, where the attacker progresses beyond easy modes toward increasingly difficult ones. We observe that this simple plug-and-play module, which seamlessly integrates into existing RL objectives, unexpectedly outperformed prior RL-based methods, improving cross-attack success rates against GFlowNets, the previous state-of-the-art, from 0.07\% to 31.28\% (a relative gain of more than 400×) with only a 6\% increase in computation.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable success in visual understanding, yet their performance degrades significantly under real-world visual corruptions. While existing robustness enhancement approaches exist, they are limited: black-box feature alignment lacks interpretability, and white-box text-based reasoning cannot restore lost pixel-level details. This work investigates a fundamental research question: Can MLLMs recover corrupted visual content by themselves? To address this, we propose Robust-U1, a novel framework that equips MLLMs with explicit visual self-recovery capability for robust understanding. The approach comprises three core stages: supervised fine-tuning for initial reconstruction, reinforcement learning with dual rewards (pixel-level SSIM and semantic-level CLIP similarity) for aligning high visual quality, and multimodal reasoning that jointly considers both the corrupted input and the recovered image. Extensive experiments demonstrate that Robust-U1 achieves state-of-the-art robustness on the real-world corruption benchmark and maintains superior performance under adversarial corruptions on general VQA benchmarks. Analysis confirms that high-quality visual recovery directly enhances reasoning performance, establishing self-recovery as a critical mechanism for robust visual understanding. The source code is available at https://github.com/jqtangust/Robust-U1.

Deploying robots in open-ended real-world environments demands continual learning capabilities to adapt to an ever-expanding range of tasks. This requires retaining previously acquired skills without forgetting while effectively leveraging prior knowledge to learn new ones. Inspired by neuroscience, we propose **N**euro-**e**volutionary **C**ontinual **R**einforcement **L**earning (**Nevo-CRL**). Nevo-CRL maintains a fixed-capacity monolithic policy network, solving tasks by optimizing inter-layer connectivity and neuron parameters. For each new task, Nevo-CRL constructs a mask population to selectively activate the outputs of each hidden layer, thereby forming a task-specific policy population. Upon completing each task, the best-performing mask is stored, and its activated neurons are frozen to prevent catastrophic forgetting. To facilitate knowledge transfer, Nevo-CRL reuses neurons from acquired skills based on semantic similarity between tasks, while dynamically allocating additional neurons for task-specific adaptation. In the learning process, Nevo-CRL iteratively adjusts masks via importance-guided crossover to optimize the policy network connectivity. To improve neuron utilization, we prune low-activity connections to recycle neurons. Experiments demonstrate that Nevo-CRL achieves state-of-the-art performance among continual RL methods. The code is available at [https://github.com/yeshenpy/Nevo-CRL](https://github.com/yeshenpy/Nevo-CRL).

Reinforcement learning (RL) often has a hierarchical structure, where an upper-level (UL) learner selects model parameters and a lower-level (LL) decision-making process responds, naturally leading to a bilevel optimization problem. Most existing bilevel RL methods assume a single-policy LL Markov decision process (MDP), and therefore fail to capture competitive structures arising in applications such as incentive design, where multiple policies interact. We study bilevel optimization problems in which the LL problem is a regularized min–max zero-sum Markov game and the UL objective is optimized through the saddle-point equilibrium induced by the LL game. In this work, we propose penalty-augmented Nikaido–Isoda descent–ascent (PANDA), a penalty-based first-order policy-gradient method based on the Nikaido–Isoda function. By exploiting the min–max game structure, PANDA avoids computing UL hypergradients and does not require second-order information. We prove that PANDA converges to stationary points without convexity assumptions on either the UL or LL objectives. Moreover, PANDA reaches an $\epsilon$-stationary point in $\tilde{\mathcal{O}}(\epsilon^{-1})$ iterations with sample complexity $\tilde{\mathcal{O}}(\epsilon^{-3})$, matching the best-known rates for bilevel RL with single-policy LL MDPs. Experiments demonstrate the superior performance of PANDA over closely related baselines.

Behavior cloning with high-capacity generative policies achieves strong imitation performance, but is often limited by demonstration coverage and distribution shift. Direct reinforcement learning fine-tuning can improve performance, but updating large action decoders is frequently unstable and sample inefficient. We propose **Lagrangian Perturbation Diffusion Steering (LP-DS)**, a lightweight adaptation method that improves a frozen generative policy by learning a compact noise-space perturbation before decoding. LP-DS optimizes this perturbation with a Lagrangian trust-region objective, improving downstream value while constraining deviation from the latent prior. Across RoboMimic manipulation, OpenAI Gym locomotion, and Adroit dexterous manipulation benchmarks, LP-DS improves sample efficiency, success, and return while maintaining higher action-space entropy than unconstrained noise-space steering, with return improvements of up to 25\% over prior baselines. Additional evaluations with flow-matching backbones, a large vision-language-action model, and physical Franka deployment show that LP-DS is not limited to compact diffusion policies or simulated benchmarks. Project page: https://sites.google.com/view/lp-ds/home.

We study a sequential resource allocation problem motivated by adaptive network recruitment, in which a limited budget of identical resources must be allocated over multiple rounds to individuals with stochastic referral capacity. Successful referrals endogenously generate future decision opportunities while allocating additional resources to an individual exhibits diminishing returns. We first show that the single-round allocation problem admits an exact greedy solution based on marginal survival probabilities. In the multi-round setting, the resulting Bellman recursion is intractable due to the stochastic, high-dimensional evolution of the frontier. To address this, we introduce a population-level surrogate value function that depends only on the remaining budget and frontier size. This surrogate enables an exact dynamic program via truncated probability generating functions, yielding a planning algorithm with polynomial complexity in the total budget. We further analyze robustness under model misspecification, proving a multi-round error bound that decomposes into a tight single-round frontier error and a population-level transition error. Finally, we evaluate our method on real-world inspired recruitment scenarios.

Decoding fMRI into natural language is challenging because strong, pre-trained language priors can dominate autoregressive generation, obscuring whether a model truly utilizes neural evidence. We introduce BIT-LLM, which exposes fMRI-derived tokens as a persistent key–value memory through interleaved cross-attention adapters, enabling repeated neural access throughout decoding. BIT-LLM is trained with a three-stage pipeline: (i) multimodal contrastive learning to obtain semantically aligned fMRI representations, (ii) supervised fine-tuning to learn the brain-LLM interface while freezing the encoder and backbone LLM, and (iii) reward-based finetuning to optimize sequence-level caption quality directly. On the NSD subject-heldout benchmark (S1-7 train, S8 test), BIT-LLM yields substantially improved captioning quality over prior baselines under greedy decoding. In addition to standard captioning metrics, we perform several complementary evaluations to assess the robustness of brain–language grounding. Specifically, we conduct perturbation-based sanity checks by zeroing fMRI inputs or shuffling voxel values, and examine whether internal representations and generated outputs change accordingly. BIT-LLM exhibits clear sensitivity to these perturbations, indicating effective utilization of voxel values and their spatial correspondence.

How does the amount of compute available to a reinforcement learning (RL) policy affect its learning? Can policies using a fixed amount of parameters, still benefit from additional compute? The standard RL framework does not provide a language to answer these questions formally. Empirically, deep RL policies are often parameterized as neural networks with static architectures, conflating the amount of compute and the number of parameters. In this paper, we formalize compute bounded policies and prove that policies which use more compute can solve problems and generalize to longer-horizon tasks that are outside the scope of policies with less compute. Building on prior work in algorithmic learning and model-free planning, we propose a minimal architecture that can use a variable amount of compute. Our experiments complement our theory. On a set 31 different tasks spanning online and offline RL, we show that $(1)$ this architecture achieves stronger performance simply by using more compute, and $(2)$ stronger generalization on longer-horizon test tasks compared to standard feedforward networks or deep residual network using upto 5 times more parameters.

Standard practice across domains from robotics to language is to first pretrain a policy on a large-scale demonstration dataset, and then finetune this policy, typically with reinforcement learning (RL), in order to improve performance on deployment domains. This finetuning step has proved critical in achieving human or super-human performance, yet while much attention has been given to developing more effective finetuning algorithms, little attention has been given to ensuring the pretrained policy is an effective initialization for RL finetuning. In this work we seek to understand how the pretrained policy affects finetuning performance, and how to pretrain policies in order to ensure they are effective initializations for finetuning. We first show theoretically that standard behavioral cloning (BC) can fail to ensure coverage over the demonstrator's actions, a minimal condition necessary for effective RL finetuning. We then show that if, instead of exactly fitting the observed demonstrations, we train a policy to model the posterior distribution of the demonstrator's behavior given the demonstration dataset, we do obtain a policy that ensures coverage over the demonstrator's actions, enabling more effective finetuning. Furthermore, this policy achieves this while ensuring pretrained performance is no worse than that of the BC policy. We then show this approach is practically implementable with modern generative models and leads to significantly improved RL finetuning performance on both realistic robotic control benchmarks and real-world robotic manipulation tasks, as compared to standard behavioral cloning.

Off-policy learning methods seek to derive an optimal policy directly from a fixed dataset of prior interactions. This objective presents significant challenges, primarily due to the inherent distributional shift and value function overestimation bias. These issues become even more noticeable in zero-shot reinforcement learning, where an agent trained on reward-free data must adapt to new tasks at test time without additional training. In this work, we address the off-policy problem in a zero-shot setting by discovering a theoretical connection of successor measures to stationary density ratios. Using this insight, our algorithm can infer optimal importance sampling ratios, effectively performing a stationary distribution correction with an optimal policy for any task on the fly. We benchmark our method in motion tracking tasks on SMPL Humanoid, continuous control on ExoRL, and for the long-horizon OGBench tasks. Our technique seamlessly integrates into forward-backward representation frameworks and enables fast-adaptation to new tasks in a training-free regime. More broadly, this work bridges off-policy learning and zero-shot adaptation, offering benefits to both research areas.

Large Language Models (LLMs) often struggle with the precise logic and schema alignment required for complex Text-to-SQL tasks. While current methods rely heavily on static prompting, they lack the ability to dynamically adapt and self-correct through environmental interaction. To bridge this gap, we propose **MARS-SQL**, a trainable multi-agent framework for Text-to-SQL. Rather than introducing a new standalone SQL primitive, **MARS-SQL** makes an agentic workflow trainable by decomposing the problem into three specialized roles: schema grounding, query generation, and solution validation. Central to our approach is a generation agent trained via a multi-turn RL policy within a ReAct-style loop. The agent learns to iteratively reason, execute intermediate SQL actions on a live database, and refine its strategy based on execution feedback. To improve robustness, we further introduce a validation mechanism that treats solution selection as a generative modeling task, identifying the optimal interaction trajectory through next-token prediction probabilities. Empirical evaluations demonstrate the effectiveness of coupling interactive learning with trajectory ranking. **MARS-SQL** achieves state-of-the-art performance, recording an execution accuracy of 77.84\% on the BIRD development dataset and 89.75\% on the Spider test dataset, while also transferring strongly to out-of-domain benchmarks. Code is available at https://github.com/YangHaolin0526/MARS-SQL.

Process rewards have been widely used in deep reinforcement learning to improve training efficiency, reduce variance, and prevent reward hacking. In LLM reasoning, existing works also explore various solutions for learning effective process reward models (PRM) with or without the help of an expert policy. However, existing methods either rely on strong assumptions about the expert policies (e.g., requiring their reward functions) or suffer intrinsic limitations (e.g., entropy collapse), resulting in weak PRMs or limited generalizability. In this paper, we introduce rePIRL, an inverse RL-inspired framework that learns effective PRMs with minimal assumptions about expert policies. Specifically, we design a dual learning process that updates the policy and the PRM interchangeably. Our learning algorithm has customized techniques to address the challenges of scaling traditional inverse RL to LLMs. We theoretically show that our proposed learning framework can unify both online and offline PRM learning methods, justifying that rePIRL can learn PRMs with minimal assumptions. Empirical evaluations on standardized math and coding reasoning datasets demonstrate the effectiveness of rePIRL over existing methods. We further show the application of our trained PRM in test-time training, test-time scaling, and providing an early signal for training hard problems. Finally, we validate our training recipe and key design choices via a detailed ablation study.

Many learning problems require uncovering a hidden ordering that reveals structure in unordered data, such as monotonicity in sorting or spatial continuity in jigsaw reconstruction. In these settings, permutations can be learned as latent operators by optimizing objectives defined directly on the reordered output, often without access to ground-truth orderings. Differentiable relaxations such as Gumbel–Sinkhorn make this approach practical by approximating permutation matrices with doubly stochastic matrices. However, learning from structure without supervision induces a non-uniform uncertainty: some assignments become confident early, while others remain ambiguous. Existing methods control this process using a single global temperature, forcing all assignments to sharpen or diffuse simultaneously and leading to instability at scale. We introduce an entropy-adaptive formulation of Gumbel–Sinkhorn that locally modulates temperature based on assignment uncertainty. This allows confident assignments to discretize early while preserving exploration where uncertainty remains. Across sorting and jigsaw reconstruction tasks and in routing-style settings, adaptive entropy control improves training stability and final permutation quality relative to fixed-temperature baselines, particularly as problem size and assignment ambiguity increase.

Few shot segmentation with vision foundation models relies on high quality prompts to segment unseen categories from limited support annotations. Existing prompt construction methods depend on test-time adaptation, fixed heuristic sampling, or monolithic reinforcement learning, making it difficult to balance semantic consistency, spatial coverage, and prompt credit assignment. To address these limitations, PromptPilot is proposed as a hierarchical multi-agent reinforcement learning framework for point prompt optimization with frozen DINOv2 and SAM. Prompt construction is formulated as sequential decision making, where feature and physical agents propose complementary prompt modifications, and a manager agent selects actions using SAM feedback and local marginal contribution. PromptPilot functions as an inference-time optimization strategy without parameter updates. Extensive experiments demonstrate that the proposed game-theoretic approach improves segmentation performance and generalization, offering a principled solution for automated prompt engineering. The code is available at [https://github.com/L-AILab/PromptPilot](https://github.com/L-AILab/PromptPilot).

Masked diffusion large language models (dLLMs) are a promising alternative to autoregressive generation. While reinforcement learning (RL) algorithms have been adapted to be compatible with dLLMs for fine-tuning them, their reliance on the computation of the marginal likelihood to evaluate policy objectives is intractable. To overcome this, we exploit a dynamical relation between the unmasking posterior of the base model and that which targets the reward-tilted distribution to derive Discrete Tilt Matching (DTM), an algorithm that avoids intractable likelihood evaluation entirely. DTM can be phrased as a cross-entropy loss that only requires forward evaluation of rewards and whose variance can be adaptively controlled, improving training stability. We motivate DTM on maze planning tasks, and show that fine-tuning LLaDA-8B-Instruct with DTM achieves higher accuracy at lower compute costs than prior RL-based fine-tuning methods across the Sudoku, Countdown, and MATH500 benchmarks.

Despite the success of LVLMs, general optimization objectives (e.g., standard MLE) fail to constrain visual trajectories, leading to language bias and hallucination. To mitigate this, current methods introduce geometric priors from visual experts as additional supervision. However, we observe that such supervision is typically suboptimal: *it is biased toward geometric precision and offers limited reasoning utility*. To bridge this gap, we propose Perceptual Flow Network (PFlowNet), which eschews rigid alignment with the expert priors and achieves interpretable yet more effective visual reasoning. Specifically, PFlowNet decouples perception from reasoning to establish a self-conditioned generation process. Based on this, it integrates *multi-dimensional rewards* with *vicinal geometric shaping* via variational reinforcement learning, thereby facilitating reasoning-oriented perceptual behaviors while preserving visual reliability. PFlowNet delivers a provable performance guarantee and competitive empirical results, particularly setting new SOTA records on V* Bench (90.6%) and MME-RealWorld-lite (67.0%).