ICML 2026 论文列表

Multi-task vehicle routing solvers via deep reinforcement learning have attracted broad attention and achieved significant progress in handling multiple constraints. However, existing neural solvers still face critical challenges, including insufficient representation, unstable training, and inefficient exploration in large combinatorial action spaces, which often prevents performance from meeting its full potential. To address these issues, we propose PoMtVRS (Preference-Optimized Multi-Task Vehicle Routing Solver with Preference Gating), a plug-and-play framework that jointly improves decoder representations and exploration efficiency through a synergistic combination of decoder-side augmentation and preference-driven optimization. Specifically, we introduce the preference optimization objective to learn relative comparisons among candidate solutions for different routing tasks, encouraging a higher generation probability of better solutions. Meanwhile, we design a preference-gated block that adaptively modulates decoder representations via sparse gated attention and nonlinear residual refinement. Extensive experiments demonstrate that PoMtVRS elevates state-of-the-art unified neural VRP backbones, achieving leading performance in multi-task benchmarks and stronger generalization.

Continual training of deep neural networks under non-stationarity often leads to a progressive loss of plasticity, eventually limiting further learning. We relate plasticity to the empirical Neural Tangent Kernel, and identify dynamical isometry (the condition that layer-wise Jacobian singular values remain close to one) as a key mechanism for preserving plasticity in continual learning. We revisit a class of networks that are almost-everywhere isometric while remaining universal Lipschitz function approximators, demonstrating that near-dynamical isometry is compatible with expressive nonlinear representations. For general architectures, we propose an efficient isometry-promoting regularization scheme and identify a novel mechanism by which it can reactivate dormant ReLU units. Building on this, we introduce AdamO, an Adam-style adaptive optimizer that decouples isometry regularization from gradient updates, analogous to AdamW. We further reinterpret prior plasticity-preserving approaches through the lens of dynamical isometry, showing that they target only a partial measure of isometry. Across supervised and reinforcement-learning continual-learning benchmarks designed to induce plasticity loss, our methods consistently match or outperform existing approaches.

As large language models (LLMs) continue to scale, it becomes increasingly challenging to grow model capacity under fixed computation budgets. We propose Path-Aligned Decompression Distillation (PADD), a framework for distilling knowledge from dense teachers without explicit routing into mixture-of-experts (MoE) students while learning high-quality routing policies. PADD organizes knowledge distillation into four stages in two phases: an initialization phase (Stage I) that builds diverse functionality in the student's experts through teacher neuron clustering and student-expert warmup, and a training phase (Stages II--IV) that integrates online adaptive distillation, path-refined policy optimization, and reward-augmented load balancing in a single training pipeline.Experiments on mathematical reasoning benchmarks demonstrate that PADD yields substantial gains over strong baselines at the same inference cost and that the MoE student can match or surpass its dense teacher. They also demonstrate effective teacher-to-student knowledge distillation and stable routing behavior.

Reinforcement learning enhances the reasoning capabilities of large language models but often involves high computational costs due to rollout-intensive optimization. Online prompt selection presents a plausible solution by prioritizing informative prompts to improve training efficiency. However, current methods either depend on costly, exact evaluations or construct prompt-specific predictive models lacking generalization across prompts. This study introduces Generalizable Predictive Prompt Selection (GPS), which performs Bayesian inference towards prompt difficulty using a lightweight generative model trained on the shared optimization history. Intermediate-difficulty prioritization and history-anchored diversity are incorporated into the batch acquisition principle to select informative prompt batches. The small predictive model also generalizes at test-time for efficient computational allocation. Experiments across varied reasoning benchmarks indicate GPS's substantial improvements in training efficiency, final performance, and test-time efficiency over superior baseline methods. The code is available at https://github.com/thu-rllab/GPS.

Low-Rank Adaptation (LoRA) is a standard tool for parameter-efficient finetuning of large models. While it induces a small memory footprint, its training dynamics can be surprisingly complex as they depend on several hyperparameters such as initialization, adapter rank, and learning rate. In particular, it is unclear *how the optimal learning rate scales with adapter rank*, which forces practitioners to re-tune the learning rate whenever the rank is changed. In this paper, we introduce *Maximal-Update Adaptation* ($\mu$A), a theoretical framework that characterizes how the "optimal" learning rate should scale with model width and adapter rank to produce stable, non-vanishing feature updates under standard configurations. Our analysis leverages techniques from hyperparameter transfer and reveals that the optimal learning rate exhibits different scaling patterns depending on initialization and LoRA scaling factor. Specifically, we identify two regimes: one where the optimal learning rate remains roughly invariant across ranks, and another where it scales inversely with rank. We further identify a configuration that allows learning rate transfer from LoRA to full finetuning, drastically reducing the cost of learning rate tuning for full finetuning. Experiments across language, vision, vision–language, image generation, and reinforcement-learning tasks validate our scaling rules and show that learning rates tuned on LoRA transfer reliably to full finetuning.

How should we analyze memory in deep RL? We introduce tools for analyzing policies under partial observability and revealing how agents use memory to make decisions. To utilize these tools, we present POPGym Arcade, a collection of Atari-inspired, hardware-accelerated environments sharing a single observation and action space. Each environment provides fully and partially observable variants, enabling counterfactual studies on observability. We find that controlled studies are necessary for fair comparisons and identify a pathology where value functions smear credit over irrelevant history. Using this pathology, we demonstrate how out-of-distribution scenarios can contaminate memory, perturbing the policy far into the future.

Offline reinforcement learning (RL) aims to optimize policies from pre-collected datasets. A bottleneck of this paradigm is managing epistemic uncertainty, which arises from limited data coverage (sample-level) and the ambiguity in identifying transition dynamics from finite data (model-level). To provide a unified quantification of these uncertainties, Bayesian RL has been proposed by treating the dynamics model as a random variable and maintaining a corresponding belief. Despite its theoretical appeal, policy optimization in Bayesian RL remains computationally challenging as it requires solving composite objectives with expectations. Prior methods either employ search-based techniques with poor computational scalability or impose restrictive posterior assumptions that sacrifice the adaptability of Bayesian RL. To address these limitations, we propose Posterior Hybrid Bayesian Belief (PhyB), which reformulates the expectation as a convex combination over a subset of dynamics models. Theoretical analysis demonstrates that the objective discrepancy induced by this approximation remains bounded. Based on PhyB, we develop an iterative regularized policy optimization algorithm that provides metric-agnostic guarantees for monotonic improvement until convergence. Empirical results demonstrate that PhyB achieves state-of-the-art performance on various benchmarks.

Reinforcement learning (RL) fine-tuning of Large Language Models (LLMs) is compute-intensive because each prompt requires generating multiple responses, or rollouts. To make the best use of GPU resources, the key question is how to allocate rollout-generation jobs across prompts. Existing methods typically use uniform allocation, assigning every prompt the same rollout budget. This is inefficient and ineffective: easy prompts are oversampled after they are already solved, while hard prompts receive too little exploration. In policy optimization methods such as Group Relative Policy Optimization (GRPO), both cases often yield near-zero gradients and limited learning progress. We address this problem by formulating rollout allocation as a computeconstrained resource allocation task, where each prompt-budget choice has an expected learning value and rollout cost. Based on this view, we propose Knapsack RL, a practical framework that uses knapsack optimization to assign heterogeneous, task-specific rollout budgets. It directs exploration toward prompts that benefit most from additional sampling. Applied to GRPO, Knapsack RL increases the effective-gradient ratio by up to 40%, enables larger budgets for challenging prompts, and improves mathematical reasoning by 2–4 points on average, with peak gains of up to 9 points. Achieving similar performance with uniform allocation requires about 2× more compute, demonstrating a simple and practical path toward scaling RL fine-tuning for LLMs.

Multimodal Large Language Models have made substantial progress on visual understanding tasks, yet they still perform poorly on high-resolution images. Prior work often attributes this limitation to perceptual constraints, arguing that MLLMs fail to recognize small objects and therefore rely on ''zoom-in" strategies to recover fine details. In contrast, our analysis shows that the dominant failure mode is background interference rather than object size. We study the "zoom-in" operation through a **hierarchical decoupling analysis** and propose the **Hierarchical Decoupling Framework**, a training-free method that turns implicit attention into explicit region selection. HiDe first performs Token-wise Attention Decoupling to disentangle question semantics and identify the most informative tokens, then uses their attention patterns to pinpoint the corresponding visual regions. It subsequently applies Layout-Preserving Decoupling to extract these regions from cluttered backgrounds and construct a compact representation that retains key spatial structure while filtering out irrelevant context. HiDe achieves state-of-the-art results on high-resolution benchmarks like Vstar Bench. It boosts Qwen2.5-VL 7B and InternVL3 8B to state of the art performance, reaching 92.1\% and 91.6\% on Vstar Bench, and even surpasses reinforcement learning based methods. After optimization, HiDe reduces memory usage by 75\% compared with the previous training-free approach. Code will be available at https://tennine2077.github.io/HiDe.github.io/.

Given a question, a language model (LM) implicitly encodes a distribution over possible answers. In practice, post-training procedures for LMs often collapse this distribution onto a single dominant mode. While this is generally not a problem for benchmark-style evaluations that assume one correct answer, many real-world tasks inherently involve multiple valid answers or irreducible uncertainty. Examples include medical diagnosis, ambiguous question answering, and settings with incomplete information. In these cases, we would like LMs to generate multiple plausible hypotheses, ideally with confidence estimates for each one, and without computationally intensive repeated sampling to generate non-modal answers. This paper describes a multi-answer reinforcement learning approach for training LMs to perform distributional reasoning over multiple answers during inference. We modify the RL objective to enable models to explicitly generate multiple candidate answers in a single forward pass, internalizing aspects of inference-time search into the model's generative process. Across question-answering, medical diagnostic, and coding benchmarks, we observe improved diversity, coverage, and set-level calibration scores compared to single answer trained baselines. Models trained with our approach require fewer tokens to generate multiple answers than competing approaches. On coding tasks, they are also substantially more accurate. These results position multi-answer RL as a principled and compute-efficient alternative to inference-time scaling procedures such as best-of-k.

Reinforcement Learning with Verifiable Reward (RLVR) is a powerful method for enhancing the reasoning abilities of Large Language Models, but its full potential is limited by a lack of exploration in two key areas: \textbf{Depth} (the difficulty of problems) and \textbf{Breadth} (the number of training instances). Our analysis of the popular GRPO algorithm reveals a bias that down-weights difficult, low-accuracy problems, which are crucial for improving reasoning skills. To address this, we introduce Difficulty Adaptive Rollout Sampling (DARS), a method that re-weights difficult problems by using targeted, multi-stage rollouts. This approach increases the number of rollout outcomes for these harder problems according to our proposed re-balancing schedules and leads to consistent gains in \textit{Pass@K}. We also found that simply enlarging the rollout size isn't effective and can even harm performance. We also investigated the role of breadth by scaling the batch size and using full-batch updates. This significantly improved \textit{Pass@1} performance by maintaining high token-level entropy, which indicates continued exploration and reduced gradient noise. Finally, we present DARS-Breadth, a combined approach that uses DARS with a large breadth of training data. This method demonstrates simultaneous gains in both \textit{Pass@K} and \textit{Pass@1}, confirming that depth (adaptive exploration) and breadth (scaling the training data) are orthogonal and essential dimensions for unlocking the full reasoning power of RLVR.

Iterative generative policies, such as diffusion models and flow matching, offer superior expressivity for continuous control but complicate Maximum Entropy Reinforcement Learning because their action log-densities are not directly accessible. To address this, we propose \textbf{Field Least-Energy Actor-Critic (FLAC)}, a likelihood-free framework that regulates policy stochasticity by penalizing the kinetic energy of the velocity field. Our key insight is to formulate policy optimization as a Generalized Schr\"odinger Bridge (GSB) problem relative to a high-entropy reference process (e.g., uniform). Under this view, the maximum-entropy principle emerges naturally as staying close to a high-entropy reference while optimizing return, without requiring explicit action densities. In this framework, kinetic energy serves as a physically grounded proxy for divergence from the reference: minimizing path-space energy bounds the deviation of the induced terminal action distribution. Building on this view, we derive an energy-regularized policy iteration scheme and a practical off-policy algorithm that automatically tunes the kinetic energy via a Lagrangian dual mechanism. Empirically, FLAC achieves superior or comparable performance on high-dimensional benchmarks relative to strong baselines, while avoiding explicit density estimation.

Reinforcement learning (RL) is a fundamental framework for sequential decision-making, in which an agent learns an optimal policy through interactions with an unknown environment. In settings with function approximation, many existing RL algorithms achieve favorable sample complexity, but often rely on computationally intractable oracles. In this paper, we use supervised learning as a computational proxy to establish a clear hierarchy of commonly adopted RL oracles under low-rank Markov Decision Processes (MDPs). This hierarchy shows that policy evaluation is the most computationally efficient oracle, provided that supervised learning can be efficiently solved. Motivated by this observation, we propose a novel optimistic actor–critic algorithm that relies solely on the policy evaluation oracle. We prove that our algorithm outperforms the existing sample complexity guarantees for low-rank MDPs while avoiding computationally expensive planning or optimization oracles commonly assumed in prior works. We further extend our theoretical results to approximately low-rank MDPs and demonstrate that this setting captures a broad class of real-world environments. Finally, we validate our theoretical results with experiments on several standard Gym benchmarks.

From a statistical evidence perspective, we establish some asymptotic optimality properties of certain multiple testing rules based on the relative belief ratio (Evans, 2015). Under the two-groups model with an additive 0-1 loss and within a Bayesian decision theoretic asymptotic framework of Bogdan et al. (2011), we show that relative belief multiple testing rules induced by a simple one-group light-tailed normal prior with a single hyperparameter achieve the same asymptotic Bayes risk as the Bayes oracle benchmark. This risk is the minimum achievable in this asymptotic framework. Despite originating from a different starting point, the evidential relative belief approach enjoys oracle properties. The relative belief multiple testing approach is fundamentally different from existing Bayesian multiple testing procedures, virtually all induced by more complex heavy-tailed one-group global-local shrinkage priors using purely posterior-based inferences (Datta & Ghosh, 2013; Ghosh et al., 2016; Bhadra et al., 2017; Ghosh & Chakrabarti, 2017; Qin & Ghosh, 2025). By measuring statistical evidence via both the prior and posterior, the relative belief approach reveals an alternative new inferential paradigm for attaining asymptotic Bayes optimality under sparsity, one that does not rely on developing increasingly elaborate priors.

Decoding-based regression, which reformulates regression as a sequence generation task, has emerged as a promising paradigm of applying large language models for numerical prediction. However, its progress is hindered by the misalignment between discrete token-level objectives (e.g., cross-entropy) and continuous numerical values. Existing approaches relying on token-level constraints often fail to capture the global magnitude of the target value, limiting their precision and generalization. In this paper, we propose to unlock the potential of decoding-based regression via reinforcement learning. We formulate the generation process as a Markov decision process, utilizing sequence-level rewards to enforce global numerical coherence.Under this framework, we present GenRe$^2$, which combines policy gradient methods and on-policy distillation to provide dense expert supervision while preserving error magnitudes, thereby resolving the temporal credit assignment challenge. Extensive experiments across tabular regression, code metric prediction and generative reward modeling demonstrate that GenRe$^2$ consistently outperforms traditional baselines, establishing a robust paradigm for general-purpose numerical prediction.

We compare the efficacy and distributional effects of supervised fine-tuning (SFT) and reinforcement learning (RL) post-training for PlasmidGPT, a foundation model for whole-plasmid generation, using Group Relative Policy Optimization (GRPO) for the RL model. Using a biologically motivated reward function encoding functional annotations, length constraints, and repeat penalties, the RL model achieves a 71.6% quality-control pass rate across 8 prompts on 4,000 sequences, compared to 4.3% for the pretrained baseline and 11.0% for SFT. A five-model reward ablation identifies the cassette arrangement bonus, which rewards correct promoter→CDS→terminator ordering, as the critical reward component. Rejectionsampling baselines indicate that the gain is not recovered by sampling more heavily from the base model. Beyond directly optimized features, RLgenerated sequences converge toward real plasmid distributions in 3-mer composition and minimum free energy density, neither of which is directly optimized by the reward function. Minimum free energy density independently converges to the real-plasmid regime under both SFT and RL despite these being parallel post-training paths. On a small curated hold-out set, RL improves continuation log-likelihood over the pretrained baseline on all 29 held-out sequences (mean ∆ = +0.83 nats).

Cooperation is central to multi-agent reinforcement learning (MARL), yet learned coordination can be fragile when external perturbations disrupt inter-agent interactions. Prior robust MARL methods have primarily considered value-oriented attacks, leaving a gap in robustness when interaction structures themselves are corrupted. In this paper, we propose an interaction-breaking adversarial learning (IBAL) framework that takes an information-theoretic view to construct attacks that impede coordination by perturbing agents’ observations and actions, and trains agents to perform reliably under such disruptions. Empirically, our approach improves robustness over existing robust MARL baselines across diverse attack settings and yields stronger performance even under agent-missing scenarios. Our code is available at https://sunwoolee0504.github.io/IBAL.

Aligning Large Language Models (LLMs) with human preferences is often formulated via Direct Preference Optimization (DPO). However, the standard Bradley-Terry instantiation of DPO is limited in modeling common departures from transitivity in human preferences. To address this, recent work has introduced Self-Play Preference Optimization (SPPO), which iteratively refines the policy by training on self-generated win-lose pairs. Our investigation, however, reveals a critical instability in SPPO: the optimization is prone to \textit{policy degeneration} when the preference oracle assigns overly confident wins to semantically indistinguishable responses. To mitigate this, we propose $\textit{S}$-SPPO, a dual-space semantic calibration framework comprising: i) $\textit{Supervision Calibration}$ via semantic gating, which anneals win rate targets toward the maximum-entropy baseline as semantic overlap increases; and ii) $\textit{Representation Calibration}$ via latent repulsion to enforce geometric diversity to prevent manifold collapse and maintain latent diversity between chosen and rejected samples. Theoretically, we show that the calibration preserves the constant-sum game structure, facilitating convergence to a Nash Equilibrium. Empirically, $\textit{S}$-SPPO avoids the performance degradation seen in prior methods, achieving 52.19\% win rate and 47.46\% length-controlled win rate on AlpacaEval 2.0 with Llama-3-8B, without using additional human-annotated preferences during training.

We introduce **Finslerian Reinforcement Learning (FiRL)**, an RL framework that makes directional costs explicit and improves robustness to tail risk. FiRL incorporates a *Finsler metric* into the locomotion cost, expressing effort as $F(x,v)$ that depends on the state $x$ and motion $v$, so it can capture uphill versus downhill asymmetry, lateral slip, and other direction-dependent effects. To handle rare but catastrophic outcomes, FiRL optimizes a Conditional Value-at-Risk ($CVaR_\alpha$) objective. We derive the corresponding risk-sensitive Bellman equation and show that the resulting CVaR--Finsler Bellman operator is a $\gamma$-contraction. This guarantees a unique fixed-point value function, while the underlying Finsler cost induces an asymmetric path cost $d_F$ that satisfies a triangle inequality despite directional asymmetry. We then develop a FiRL actor--critic algorithm to learn policies under this anisotropic, risk-averse objective. Across simulation benchmarks and real-world robot trials, FiRL demonstrates safer and more energy-efficient locomotion behavior than strong baselines such as risk-neutral PPO. For instance, on a $12^\circ$ sloped Hopper task, FiRL reduces worst-case ($CVaR_{0.1}$) impact forces by over 35% and total energy cost by 15%, while also improving success rate.