ICML 2026 Papers

Equivariant neural networks exploit underlying task symmetries to improve generalization, but strict equivariance constraints can induce more complex optimization dynamics that can hinder learning. Prior work addresses these limitations by relaxing strict equivariance during training, but typically relies on prespecified, explicit, or implicit target levels of relaxation for each network layer, which are task-dependent and costly to tune. We propose Recurrent Equivariant Constraint Modulation (RECM), a layer-wise constraint modulation mechanism that learns appropriate relaxation levels solely from the training signal and the symmetry properties of each layer's input-target distribution, without requiring any prior knowledge about the task-dependent target relaxation level. We demonstrate that under the proposed RECM update, the relaxation level of each layer provably converges to a value upper-bounded by its symmetry gap, namely the degree to which its input-target distribution deviates from exact symmetry. Consequently, layers processing symmetric distributions recover full equivariance, while those with approximate symmetries retain sufficient flexibility to learn non-symmetric solutions when warranted by the data. Empirically, RECM outperforms prior methods across diverse exact and approximate equivariant tasks, including the challenging molecular conformer generation on the GEOM-Drugs dataset.

Egocentric reasoning fundamentally differs from third-person understanding in LVLMs. Third-person settings offer wide and stable contexts with consistent global regularities, allowing models to utilize broad statistical correlations. In contrast, egocentric scenes are highly dynamic and heterogeneous, where decisive cues are localized and atypical. Therefore, robust egocentric reasoning requires models to focus on ''what is seen now'', i.e., the immediate visual input. However, existing methods tend to exhibit "inertial thinking'', relying excessively on language priors and global context. To address this limitation, we propose a novel three-stage Ego3S framework to ground models' reasoning in interaction evidence. Specifically, before training, we first utilize the counterfactual-based paradigm to select high-value samples that effectively activate multimodal reasoning, thus mitigating the over-reliance on language priors and global context. Moreover, we introduce an interaction-centric reward for reinforcement learning that strengthens the model’s sensitivity to localized interaction cues. Finally, during training, we employ a variance-aware learning schedule that monitors reward distributions to dynamically synchronize data selection with the evolving model competence. Experiments on five datasets show that our Ego3S consistently achieves superior performance using only 26.5% of the training data, while reducing computational costs by over 46%. Code is available at https://github.com/Mrshenshen/Ego3S.

Large language model (LLM) watermarking has shown promise in detecting AI-generated content and mitigating misuse, with prior work claiming robustness against paraphrasing and text editing. In this paper, we argue that existing evaluations are not sufficiently adversarial, obscuring critical vulnerabilities and overstating the security. To address this, we introduce the *adaptive robustness radius*, a formal metric that quantifies the worst-case resilience of watermarks against adaptive adversaries. By lifting the paraphrase space into a KL-divergence ball, we approximate this radius and theoretically demonstrate that optimizing the attack context and model parameters can significantly reduce the approximated radius, making watermarks highly vulnerable to paraphrase attacks. Leveraging this insight, we propose RLCracker, a reinforcement learning (RL)–based adaptive attack that erases watermarks while preserving semantic fidelity. RLCracker requires only *limited* watermarked examples and *zero* access to the detector. Despite weak supervision, it empowers a 3B model to achieve 98.5\% removal success with minimal semantic shift on 1,500-token Unigram-marked texts after training on only *100* short samples. This performance dramatically exceeds 6.75\% by GPT-4o and generalizes across five model sizes over ten watermarking schemes. Our code is available in this repository.

Neural tangent kernels (NTKs) are a powerful tool for analyzing deep, non-linear neural networks. In the infinite-width limit, NTKs can easily be computed for most common architectures, yielding full analytic control over the training dynamics. However, at infinite width, important properties of training such as NTK evolution or feature learning are absent. Nevertheless, finite width effects can be included by computing corrections to the Gaussian statistics at infinite width. We introduce Feynman diagrams for computing finite-width corrections to NTK statistics. These dramatically simplify the necessary algebraic manipulations and enable the computation of layer-wise recursion relations for arbitrary statistics involving preactivations, NTKs and certain higher-derivative tensors (dNTK and ddNTK) required to predict the training dynamics at leading order. We demonstrate the feasibility of our framework by extending stability results for deep networks from preactivations to NTKs and proving the absence of finite-width corrections for scale-invariant nonlinearities such as ReLU on the diagonal of the Gram matrix of the NTK. We numerically implement the complete set of equations necessary to compute the first-order corrections for arbitrary inputs and demonstrate that the results follow the statistics of sampled neural networks for widths $n\gtrsim 20$.

Learning domain adaptive policies that can generalize to unseen transition dynamics, remains a fundamental challenge in learning-based control. Substantial progress has been made through domain representation learning to capture domain-specific information, thus enabling domain-aware decision making. We analyze the process of learning domain representations through dynamical prediction and find that selecting contexts adjacent to the current step causes the learned representations to entangle static domain information with varying dynamical properties. Such mixture can confuse the conditioned policy, thereby constraining zero-shot adaptation. To tackle the challenge, we propose DADP (Domain Adaptive Diffusion Policy), which achieves robust adaptation through unsupervised disentanglement and domain-aware diffusion injection. First, we introduce Lagged Context Dynamical Prediction, a strategy that conditions future state estimation on a historical offset context; by increasing this temporal gap, we unsupervisedly disentangle static domain representations by filtering out transient properties. Second, we integrate the learned domain representations directly into the generative process by biasing the prior distribution and reformulating the diffusion target. Extensive experiments on challenging benchmarks across locomotion and manipulation demonstrate the superior performance, and the generalizability of DADP over prior methods. More visualization results are available on the https://outsider86.github.io/DomainAdaptiveDiffusionPolicy/.

Tool-integrated (TI) reinforcement learning (RL) enables large language models (LLMs) to perform multi-step reasoning by interacting with external tools such as search engines and retrievers. Group Relative Policy Optimization (GRPO), exemplified by the recent Search-R1, offers fast convergence and a value-free formulation that makes it appealing for this setting, yet consistently suffers from training collapse. We identify Lazy Likelihood Displacement (LLD), a systematic reduction or stagnation in the likelihood of both correct and incorrect responses, as the core mechanism driving this failure. LLD emerges early and triggers a self-reinforcing LLD Death Spiral, where declining likelihood leads to low-confidence responses, inflating gradients, and ultimately causing collapse. We empirically characterize this process across models on a Search-R1-style, search-integrated question answering task, revealing a consistent three-phase trajectory: early stagnation, steady decay, and accelerated collapse. To address this, we propose a likelihood-preserving regularization LLDS that activates only when a response action’s likelihood decreases, and regularizes only the tokens responsible. This fine-grained structure mitigates LLD with minimal interference. Our method stabilizes training, prevents gradient explosion, and yields substantial performance improvements across seven benchmarks, including relative improvements of +45.2% on Qwen2.5-3B and +37.1% on Qwen2.5-7B over vanilla GRPO training. Our results establish LLD as a previously overlooked bottleneck in GRPO- based TIRL and provide a practical path toward stable, scalable training of tool-integrated RL.

Diffusion Large Language Models (dLLMs) introduce a new paradigm for language generation, which in turn presents new challenges for aligning them with human preferences. In this work, we aim to improve the policy optimization for dLLMs by reducing the cost of the trajectory probability calculation, thereby enabling scaled-up offline policy training. We prove that: (i) under reference policy regularization, the probability ratio of the newly unmasked tokens is an unbiased estimate of that of intermediate diffusion states, and (ii) the probability of the full trajectory can be effectively estimated with a single forward pass of a re-masked final state. By integrating these two trajectory reduction strategies into a policy optimization objective, we propose Trajectory Reduction Policy Optimization (dTRPO). We evaluate dTRPO on 7B dLLMs across instruction-following and reasoning benchmarks. Results show that it substantially improves the core performance of state-of-the-art dLLMs, achieving gains of up to 9.6% on STEM tasks, up to 4.3% on coding tasks, and up to 3.0% on instruction-following tasks. Moreover, dTRPO exhibits strong training efficiency due to its offline, single-forward nature, and achieves improved generation efficiency through high-quality outputs.

Reinforcement Learning with Verifiable Rewards (RLVR) is highly effective for enhancing LLM reasoning, yet recent evidence shows models like Qwen2.5 achieve significant gains even with spurious rewards. We investigate this phenomenon and identify ``Perplexity Paradox'': spurious RLVR triggers a divergence where answer-token perplexity drops while prompt-side coherence degrades, suggesting model is bypassing reasoning in favor of memorization. Using a suite of mechanistic interpretability tools, including Path Patching and Logit Lens, we identify a previously unknown Anchor–Adapter circuit. This circuit enables model to bypass reasoning and directly retrieve memorized solutions under spurious RLVR. We localize a Functional Anchor in middle layers (L18–20) that triggers retrieval of memorized solutions, followed by Structural Adapters in later layers (L21+) that transform representations to accommodate shortcut signal. Finally, we demonstrate that scaling specific MLP keys within this circuit allows for bidirectional causal steering, i.e., artificially amplifying or suppressing contamination-driven performance. Our results provide a mechanistic roadmap for identifying and mitigating data contamination in RLVR-tuned models.

Recently, small models with latent recursion have obtained promising results on complex reasoning tasks. These results are typically explained by the theory that such recursion increases a network’s depth, allowing it to compactly emulate the capacity of larger models. However, the performance of recursively added layers remains behind the capabilities of one‑pass models with the same feed-forward depth. This means that in the looped version, not every recursive step effectively contributes to depth. This raises the question: when and why does latent reasoning improve performance, and when does it result in dead compute? In our work, we analyze the algorithms that latent reasoning provides answer to this question. We show that latent reasoning can be formalized as a classifier‑free guidance and policy improvement algorithm. Building on these insights, we propose to use a training schemes from RL and diffusion methods for latent reasoning modles. Using the Tiny Recursive Model as our testbed, we show that with our modifications we can avoid dead compute steps and reduce the total number of forward passes by 18× while maintaining performance. Broadly speaking, we show how a policy improvement perspective on recursive steps can explain model behavior and provide insights for further improvements.

How can we train agents to navigate uncertainty over long horizons? In this work, we propose ∆Belief-RL, which leverages a language model's own intrinsic beliefs to reward intermediate progress. Our method utilizes the change in the probability an agent assigns to the target solution for credit assignment. By training on synthetic interaction data, ∆Belief-RL teaches information-seeking capabilities that consistently outperform purely outcome-based rewards for RL, with improvements generalizing to out-of-distribution applications ranging from customer service to personalization. Notably, the performance continues to improve as we scale test-time interactions beyond the training horizon, with interaction-efficiency increasing even on Pass@k metrics. Overall, our work introduces a scalable training strategy for navigating uncertainty over a long-horizon, by enabling credit assignment to intermediate actions via intrinsic ∆Belief rewards.

This paper deals with stochastic optimization problems involving Markovian noise with a zero-order oracle. We present and analyze a novel derivative-free method for solving such problems in strongly convex smooth and non-smooth settings with both one-point and two-point feedback oracles. Using a randomized batching scheme, we show that when mixing time $\tau$ of the underlying noise sequence is less than the dimension of the problem $d$, the convergence estimates of our method do not depend on $\tau$. This observation provides an efficient way to interact with Markovian stochasticity: instead of invoking the expensive first-order oracle, one should use the zero-order oracle. Finally, we complement our upper bounds with the corresponding lower bounds. This confirms the optimality of our results.

Diffusion models have become a dominant paradigm for high-quality generative modeling, while post-training is essential for adapting them to diverse downstream applications. However, post-training of large diffusion models is still challenging due to the prohibitive memory footprints and slow training speed, which existing parameter-efficient fine-tuning methods only partially address. To overcome these limitations, we propose FourTune, an efficient post-training framework for diffusion models based on an end-to-end W4A4G4 paradigm. FourTune introduces a triple-branch hybrid pipeline that augments the standard LoRA architecture with a frozen numerical stabilizer to isolate quantization-sensitive outliers, enabling stable training under native 4-bit computation. In addition, FourTune employs hardware-efficient block-wise quantization and customized fused kernels to support efficient quantized backpropagation and reduce memory bandwidth overhead. Across customization, reinforcement learning, and distillation tasks, FourTune matches the quality of full-precision fine-tuning. On FLUX.1-dev (12B), FourTune reduces memory overhead by $2.25\times$ and increases end-to-end training throughput by $2.27\times$ compared to BF16 LoRA.

This paper proposes a suite of rationality measures and associated theory for reinforcement learning agents, a property increasingly critical yet rarely explored. We define an action in deployment to be perfectly rational if it maximises the hidden true value function in the steepest direction. The expected value discrepancy of a policy's actions against their rational counterparts, culminating over the trajectory in deployment, is defined to be expected rational risk; an empirical average version in training is also defined. Their difference, termed as rational risk gap, is decomposed into (1) an extrinsic component caused by environment shifts between training and deployment, and (2) an intrinsic one due to the algorithm's generalisability in a dynamic environment. They are upper bounded by, respectively, (1) the $1$-Wasserstein distance between transition kernels and initial state distributions in training and deployment, and (2) the empirical Rademacher complexity of the value function class. Our theory suggests hypotheses on the benefits from regularisers (including layer normalisation, $\ell_2$ regularisation, and weight normalisation) and domain randomisation, as well as the harm from environment shifts. Experiments are in full agreement with these hypotheses. The code is available at \href{https://github.com/EVIEHub/Rationality}{https://github.com/EVIEHub/Rationality}.

We study stochastic minimum-cost reach-avoid reinforcement learning, where an agent must satisfy a reach-avoid specification with probability at least $p$ while minimizing expected cumulative costs in stochastic environments. Existing safe and constrained reinforcement learning methods typically fail to jointly enforce probabilistic reach-avoid constraints and optimize cost in the learning setting in stochastic environments. To address this challenge, we introduce reach-avoid probability certificates (RAPCs), which identify states from which stochastic reach-avoid constraints are satisfiable. Building on RAPCs, we develop a contraction-based Bellman formulation that serves as a principled surrogate for integrating reach-avoid considerations into reinforcement learning, enabling cost optimization under probabilistic constraints. We establish almost sure convergence of the proposed algorithms to locally optimal policies with respect to the resulting objective. Experiments in the MuJoCo simulator demonstrate improved cost performance and consistently higher reach-avoid satisfaction rates.

Flow Matching shows great promise in offline reinforcement learning (RL), yet optimizing these iterative policies via Backpropagation Through Time (BPTT) is unstable. While prevailing paradigms circumvent this by distilling multi-step flows into single-step approximations, such methods may limit the benefits of iterative refinement. To avoid these sacrifices, we propose Direct Flow Q-Learning (DFQL), a streamlined framework that attains superior results by optimizing flow matching policies without BPTT or distillation. DFQL derives a surrogate objective that directly injects terminal Q-value gradients as a guidance term into each step velocity field, ensuring stable optimization while preserving iterative expressive capacity. Across 73 challenging tasks in OGBench and D4RL, DFQL achieves state-of-the-art results. Additionally, DFQL extends seamlessly to the offline-to-online setting, delivering substantial performance gains without further modification.

While most AI alignment research focuses on preventing models from generating explicitly harmful content, a more subtle risk arises from capability-seeking RL training in vulnerable environments. We investigate whether language models, when trained with reinforcement learning (RL) in environments with implicit loopholes, can learn to exploit these flaws to maximize reward, even without being explicitly instructed to do so. To test this, we design a suite of four diverse "vulnerability games'', each presenting a structural vulnerability related to context-conditional compliance, proxy metrics, reward tampering, and self-evaluation. Our experiments show that models often learn to exploit these vulnerabilities, discovering opportunistic strategies that increase reward while sometimes preserving or even improving standard task-performance metrics. More critically, we find that these exploitative strategies are not always narrow "tricks'': they can transfer in structured but limited ways, propagate from a capable teacher model to other student models through SFT, and in several cases remain more persistent when learned through RL than when distilled through SFT. Our findings show that alignment risks from capability-seeking RL training can be difficult to detect with standard performance monitoring, suggesting that future AI safety work should extend beyond content moderation to auditing and securing training environments, reward mechanisms, and evaluation channels. Code is available at https://github.com/YujunZhou/Capability-seeking-RL-risk.

Reinforcement learning with verifiable rewards (RLVR) has emerged as an effective post-training paradigm for improving the reasoning capabilities of large language models. However, existing group-based RLVR methods often suffer from severe sample inefficiency. This inefficiency stems from reliance on point estimation of rewards from a small number of rollouts, leading to high estimation variance, variance collapse, and ineffective utilization of generated responses. In this work, we reformulate RLVR from a statistical estimation perspective by modeling rewards as samples drawn from a policy-induced distribution and casting advantage computation as the problem of estimating the reward distribution from finite data. Building on this view, we propose **D**iscounted **B**eta–**B**ernoulli (**DBB**) reward estimation, which leverages historical reward statistics for the non-stationary distribution. Although biased, the resulting estimator exhibits reduced and stable variance and theoretically avoids variance collapse. Under mild non-stationarity, it also achieves a lower mean squared error than standard point estimation, as we characterize analytically and verify empirically. Across six in-distribution and three out-of-distribution reasoning benchmarks, GRPO with DBB consistently outperforms naive GRPO and strong recent baselines, including the replay-based RePO and the variance-collapse-aware GRESO and DAPO. Relative to GRPO, it achieves average Acc@8 improvements of 3.43/2.32 points in-distribution and 10.05/8.34 points out-of-distribution on the 1.7B and 8B models, respectively, without additional computational cost or memory usage.

We analyze the Bayesian regret of the Gaussian process posterior sampling reinforcement learning (GP-PSRL) algorithm. Posterior sampling is a heuristic for decision-making under uncertainty that has been used to develop successful algorithms for a variety of continuous control problems. However, theoretical work on GP-PSRL is limited. All known regret bounds either have a sub-optimal growth rate, require strong smoothness assumptions, or fail to properly account for the fact that the set of possible system states is unbounded. Through a recursive application of the Borell-Tsirelson-Ibragimov-Sudakov inequality, we show that, with high probability, the states actually visited by the algorithm are contained within a ball of near-constant radius. We then use the chaining method to control the regret suffered by GP-PSRL under weak smoothness conditions. Our main result is a Bayesian regret bound of the order $\widetilde{\mathcal{O}}(H\sqrt{\gamma_TT})$, where $H$ is the horizon, $T$ is the number of time steps and $\gamma_T$ is the expected information gain. With this result, we resolve the limitations with prior theoretical work on PSRL, and provide the theoretical foundation and tools for analyzing PSRL in complex settings.

Training LLMs as interactive agents for multi-turn decision-making remains challenging, particularly in long-horizon tasks with sparse and delayed rewards, where agents must execute extended sequences of actions before receiving meaningful feedback. Most existing reinforcement learning (RL) methods model LLM agents as *flat* policies operating at a single time scale, selecting an action at each turn. In sparse-reward settings, this forces the agent to infer long-range dependencies solely from distant end-of-trajectory signals, often leading to inefficient learning and unstable behavior in complex environments. We propose **HiPER**, a novel **Hi**erarchical **P**lan–**E**xecute **R**L framework that jointly models and optimizes high-level subgoal planning and low-level action execution for LLM agents to overcome flat RL's brittle long-horizon behavior and weak credit assignment under sparse outcome feedback. By maintaining persistent subgoals across multiple turns and explicitly deciding when to switch between them, HiPER introduces structured intermediate decision points that facilitate learning under sparse feedback, converting implicit multi-turn structure into learnable decisions at different time scales. To enable effective training, we introduce Hierarchical Advantage Estimation (HAE), a two-timescale policy gradient method that assigns credit to both action execution and subgoal transitions and achieves variance reduction relative to flat advantage estimation. Empirically, HiPER achieves state-of-the-art performance on challenging interactive benchmarks, reaching 97.4\% success on ALFWorld (+6.6\% over the best prior method) and 83.3\% on WebShop, with especially large gains on long-horizon tasks requiring multiple dependent subtasks. These results highlight the importance of explicit hierarchical decomposition for scalable RL training of multi-turn LLM agents.