ICLR 2026 Papers

Offline Safe Reinforcement Learning (OSRL) aims to learn a policy that achieves high performance in sequential decision-making while satisfying safety constraints, using only pre-collected datasets. Recent works, inspired by the strong capabilities of Generative Models (GMs), reformulate decision-making in OSRL as a conditional generative process, where GMs generate desirable actions conditioned on predefined reward and cost return-to-go values. However, GM-assisted methods face two major challenges in constrained settings: (1) they lack the ability to ``stitch'' optimal transitions from suboptimal trajectories within the dataset, and (2) they struggle to balance reward maximization with constraint satisfaction, particularly when tested with imbalanced human-specified reward-cost conditions. To address these issues, we propose Goal-Assisted Stitching (GAS), a novel algorithm designed to enhance stitching capabilities while effectively balancing reward maximization and constraint satisfaction. To enhance the stitching ability, GAS first augments and relabels the dataset at the transition level, enabling the construction of high-quality trajectories from suboptimal ones. GAS also introduces novel goal functions, which estimate the optimal achievable reward and cost goals from the dataset. These goal functions, trained using expectile regression on the relabeled and augmented dataset, allow GAS to accommodate a broader range of reward-cost return pairs and achieve a better tradeoff between reward maximization and constraint satisfaction compared to human-specified values. The estimated goals then guide policy training, ensuring robust performance under constrained settings. Furthermore, to improve training stability and efficiency, we reshape the dataset to achieve a more uniform reward-cost return distribution. Empirical results validate the effectiveness of GAS, demonstrating superior performance in balancing reward maximization and constraint satisfaction compared to existing methods.

From a first-principles perspective, it may seem odd that the strongest results in foundation model fine-tuning (FT) are achieved via a relatively complex, two-stage training procedure. Specifically, one first trains a reward model (RM) on some dataset (e.g., human preferences) before using it to provide *online* feedback as part of a downstream reinforcement learning (RL) procedure, rather than directly optimizing the policy parameters on said dataset via *offline* maximum likelihood estimation. In fact, from an information-theoretic perspective, we can only *lose* information via passing through a reward model and cannot create any new information via on-policy sampling. To explain this discrepancy, we scrutinize several hypotheses on the value of RL in FT through both theoretical and empirical lenses. Of the hypotheses considered, we find the most support for the explanation that on problems with a *generation-verification gap*, *(1)* it is relatively easy to learn the relatively simple RM (*verifier*) from the preference data. Then, *(2)* the downstream RL procedure only returns policies (*generators*) that are optimal for such relatively simple verifiers. Thus, end-to-end, two-stage online FT only has to search over a reduced subset of the full space of policies, requiring less data than offline FT.

Neural approaches to the Flexible Job Shop Scheduling Problem (FJSP), particularly those based on deep reinforcement learning (DRL), have gained growing attention in recent years. Yet existing methods often rely on cumbersome state representations (i.e. sometimes requiring more than 20 handcrafted features) and suboptimal neural architectures. We introduce \textsc{ReSched}, a minimalist DRL framework that rethinks both the scheduling formulation and model design. First, we revisit the Markov Decision Process (MDP) formulation of FJSP, reducing the state to just four essential features and replacing historical dependencies with a graph structure that directly encodes intra-job operation relationships. Second, we employ Transformer blocks with dot-product attention, augmented by three lightweight but effective architectural modifications tailored to scheduling. Extensive experiments show that \textsc{ReSched} outperforms classical dispatching rules and state-of-the-art DRL methods on FJSP. Moreover, \textsc{ReSched} generalizes well to the Job Shop Scheduling Problem (JSSP) and the Flexible Flow Shop Scheduling Problem (FFSP), achieving competitive performance against neural baselines specifically designed for these variants.

Ensemble learning is a practical family of methods for uncertainty modeling, particularly useful for sequential decision-making problems like recommendation systems and reinforcement learning tasks. The posterior on likelihood parameters is approximated by sampling an ensemble member from a predetermined index distribution, with the ensemble’s diversity reflecting the degree of uncertainty. In this paper, we propose Bayesian Ensemble (BE), a lightweight yet principled Bayesian layer atop existing ensembles. BE treats the selection of an ensemble member as a bandit problem in itself, dynamically updating a sampling distribution over members via Bayesian inference on observed rewards. This contrasts with prior works that rely on fixed, uniform sampling. We extend this framework to both bandit learning and reinforcement learning, introducing Bayesian Ensemble Bandit and Bayesian Ensemble Deep Q-Network for diverse decision-making problems. Extensive experiments on both synthetic and real-world environments demonstrate the effectiveness and efficiency of BE.

Estimating the state of an environment from high-dimensional, multimodal, and noisy observations is a fundamental challenge in reinforcement learning (RL). Traditional approaches rely on probabilistic models to account for the uncertainty, but often require explicit noise assumptions, in turn limiting generalization. In this work, we contribute a novel method to learn a structured latent representation, in which distances between states directly correlate with the minimum number of actions required to transition between them. The proposed metric space formulation provides a geometric interpretation of uncertainty without the need for explicit probabilistic modeling. To achieve this, we introduce a multimodal latent transition model and a sensor fusion mechanism based on inverse distance weighting, allowing for the adaptive integration of multiple sensor modalities without prior knowledge of noise distributions. We empirically validate the approach on a range of multimodal RL tasks, demonstrating improved robustness to sensor noise and superior state estimation compared to baseline methods. Our experiments show enhanced performance of an RL agent via the learned representation, eliminating the need of explicit noise augmentation. The presented results suggest that leveraging transition-aware metric spaces provides a principled and scalable solution for robust state estimation in sequential decision-making.

Distilling robust reasoning capabilities from large language models (LLMs) into smaller, computationally efficient student models remains an unresolved challenge. Despite recent advances, distilled models frequently suffer from superficial pattern memorization and subpar generalization. To overcome these limitations, we introduce a novel distillation framework that moves beyond simple mimicry to instill a deeper conceptual understanding. Our framework features two key innovations. \underline{\textit{First}}, to address pattern memorization, Explanatory Inversion (EI) generates targeted ``explanatory probes'' that compel the student to articulate the underlying logic behind an answer, rather than just memorizing it. \underline{\textit{Second}}, to improve generalization, Explanatory GRPO (\texttt{EXGRPO}) uses a reinforcement learning algorithm with a novel Dialogue Structure Utility Bonus, which explicitly rewards the student for maintaining a coherent reasoning process across these probes. Extensive evaluations on 12 datasets demonstrate significant improvements. Using Gemma-7b as the student model, our method yields an average \textbf{20.39\%} increase over zero-shot performance and a \textbf{6.02\%} improvement over the state-of-the-art distillation baselines. Moreover, models distilled with our method show remarkable training efficiency (e.g., surpassing vanilla fine-tuning with \textbf{10-25\%} training data) and strong generalization to out-of-distribution tasks.

We introduce a sequence-conditioned critic for Soft Actor--Critic (SAC) that models trajectory context with a lightweight Transformer and trains on aggregated $N$-step targets. Unlike prior approaches that (i) score state--action pairs in isolation or (ii) rely on actor-side action chunking to handle long horizons, our method strengthens the critic itself by conditioning on short trajectory segments and integrating multi-step returns---without importance sampling (IS). The resulting sequence-aware value estimates capture temporal structure critical for extended-horizon and sparse-reward problems. On local-motion benchmarks, we further show that freezing critic parameters for several steps makes our update compatible with CrossQ's core idea, enabling stable training without a target network. Despite its simplicity---a 2-layer Transformer with 128--256 hidden units and a maximum update-to-data ratio (UTD) of $1$---the approach consistently outperforms standard SAC and strong off-policy baselines, with particularly large gains on long-trajectory control. These results highlight the value of sequence modeling and $N$-step bootstrapping on the critic side for long-horizon reinforcement learning.

Neural solvers have achieved impressive progress in addressing simple routing problems, particularly excelling in computational efficiency. However, their advantages under complex constraints remain nascent, for which current constraint-handling schemes via feasibility masking or implicit feasibility awareness can be inefficient or inapplicable for hard constraints. In this paper, we present Construct-and-Refine (CaR), the first general and efficient constraint-handling framework for neural routing solvers based on explicit learning-based feasibility refinement. Unlike prior construction-search hybrids that target reducing optimality gaps through heavy improvements yet still struggle with hard constraints, CaR achieves efficient constraint handling by designing a joint training framework that guides the construction module to generate diverse and high-quality solutions well-suited for a lightweight improvement process, e.g., 10 steps versus 5k steps in prior work. Moreover, CaR presents the first use of construction-improvement-shared representation, enabling potential knowledge sharing across paradigms by unifying the encoder, especially in more complex constrained scenarios. We evaluate CaR on typical hard routing constraints to showcase its broader applicability. Results demonstrate that CaR achieves superior feasibility, solution quality, and efficiency compared to both classical and neural state-of-the-art solvers. Our code, pre-trained models, and datasets are available at: https://github.com/jieyibi/CaR-constraint.

DPO and related algorithms align language models by directly optimizing the RLHF objective: find a policy that maximizes the Bradley-Terry reward while staying close to a reference policy through a KL divergence penalty. Previous work showed that this approach could be further generalized: the original problem remains tractable even if the KL divergence is replaced by a family of $f$-divergence with a convex generating function $f$. Our first contribution is to show that convexity of $f$ is not essential. Instead, we identify a more general condition, referred to as DPO-inducing, that precisely characterizes when the RLHF problem remains tractable. Our next contribution is to establish a second condition on $f$ that is necessary to prevent probability displacement, a known empirical phenomenon in which the probabilities of the winner and the loser responses approach zero. We refer to any $f$ that satisfies this condition as displacement-resistant. We finally focus on a specific DPO-inducing and displacement-resistant $f$, leading to our novel SquaredPO loss. Compared to DPO, this new loss offers stronger theoretical guarantees while performing competitively in practice.

Reinforcement learning in combinatorial action spaces requires searching over exponentially many joint actions to simultaneously select multiple sub-actions that form coherent combinations. Existing approaches either simplify policy learning by assuming independence across sub-actions, which often yields incoherent or invalid actions when coordination is required, or attempt to learn action structure and control jointly, which is slow and unstable. We introduce Structured Policy Initialization (SPIN), a two-stage framework that first pre-trains an Action Structure Model (ASM) to capture the manifold of valid actions, then freezes this representation and trains lightweight policy heads for control. On challenging DM Control benchmarks, SPIN improves average return by up to $39\%$ over the state of the art while reducing time to convergence by up to $12.8\times$.

The use of target networks is a popular approach for estimating value functions in deep Reinforcement Learning (RL). While effective, the target network remains a compromise solution that preserves stability at the cost of slowly moving targets, thus delaying learning. Conversely, using the online network as a bootstrapped target is intuitively appealing, albeit well-known to lead to unstable learning. In this work, we aim to obtain the best out of both worlds by introducing a novel update rule that computes the target using the **MIN**imum estimate between the **T**arget and **O**nline network, giving rise to our method, **MINTO**. Through this simple, yet effective modification, we show that MINTO enables faster and stable value function learning, by mitigating the potential overestimation bias of using the online network for bootstrapping. Notably, MINTO can be seamlessly integrated into a wide range of value-based and actor-critic algorithms with a negligible cost. We evaluate MINTO extensively across diverse benchmarks, spanning online and offline RL, as well as discrete and continuous action spaces. Across all benchmarks, MINTO consistently improves performance, demonstrating its broad applicability and effectiveness.

Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning capabilities of large language models. While existing analyses identify that RLVR-induced changes are sparse, they primarily focus on the **magnitude** of these updates, largely overlooking their **direction**. In this work, we argue that the direction of updates is a more critical lens for understanding RLVR's effects, which can be captured by the signed, token-level log probability difference $\Delta\log p$ between the base and final RLVR models. Through statistical analysis and token-replacement interventions, we demonstrate that $\Delta\log p$ more effectively identifies sparse, yet reasoning-critical updates than magnitude-based metrics (e.g., divergence or entropy). Building on this insight, we propose two practical applications: (1) a *test-time extrapolation* method that amplifies the policy along the learned $\Delta\log p$ direction to improve reasoning accuracy without further training; (2) a *training-time reweighting* method that focuses learning on low-probability (corresponding to higher $\Delta\log p$) tokens, which improves reasoning performance across models and benchmarks. Our work establishes the direction of change as a key principle for analyzing and improving RLVR.

While deep reinforcement learning (DRL) from pixels has achieved remarkable success, its sample inefficiency remains a critical limitation for real-world applications. Model-based RL (MBRL) addresses this by learning a world model to generate simulated experience, but standard approaches that rely on pixel-level reconstruction losses often fail to capture small, task-critical objects in complex, dynamic scenes. We posit that an object-centric representation can direct model capacity toward semantically meaningful entities, improving dynamics prediction and sample efficiency. In this work, we introduce **OC-STORM**, an object-centric MBRL framework that enhances a learned world model with object representations extracted by a pretrained segmentation network. By conditioning on a minimal number of annotated frames, OC-STORM learns to track decision‐relevant object dynamics and inter‑object interactions without extensive labeling or access to privileged information. Empirical results demonstrate that OC-STORM significantly outperforms the STORM baseline on the Atari 100k benchmark and achieves state-of-the-art sample efficiency on challenging boss fights in the visually complex game **Hollow Knight**. Our findings underscore the potential of integrating object-centric priors into MBRL for complex visual domains. Core code and evaluation videos are available in supplementary materials.

Policy gradient algorithms have been successfully applied to enhance the reasoning capabilities of large language models (LLMs). KL regularization is ubiquitous, yet the design surface, choice of KL direction (forward vs. reverse), normalization (normalized vs. unnormalized), and estimator ($k_1/k_2/k_3$), is scattered across the literature and often intertwined with off-policy estimation. We ask a focused question: under the off-policy setting, what weighting is required for each KL variant so that the surrogate we optimize yields the exact gradient of the intended KL-regularized objective? We answer this with a compact, unified derivation we call the Regularized Policy Gradient (\textbf{RPG}) view. RPG (i) unifies normalized and unnormalized KL variants and shows that the widely-used $k_3$ penalty is exactly the unnormalized KL; (ii) specifies conditions under which REINFORCE-style losses with stop-gradient are gradient-equivalent to fully differentiable surrogates; (iii) identifies and corrects an off-policy importance-weighting mismatch in GRPO's KL term; and (iv) introduces RPG-Style Clip, a truncated-importance-sampling step within RPG-REINFORCE that enables stable, off-policy policy-gradient training at scale. On mathematical reasoning benchmarks (AIME24, AIME25), RPG-REINFORCE with RPG-Style Clip improves accuracy by up to $+6$ absolute percentage points over DAPO. Notably, RPG is a \emph{stable and scalable} RL algorithm for LLM reasoning, realized via (a) a KL-correct objective, (b) truncated importance sampling, and (c) an iterative reference-policy update scheme.

A fundamental challenge in multi-task reinforcement learning (MTRL) is achieving sample efficiency in visual domains where tasks exhibit significant heterogeneity in both observations and dynamics. Model-based RL (MBRL) offers a promising path to sample efficiency through world models, but standard monolithic architectures struggle to capture diverse task dynamics, leading to poor reconstruction and prediction accuracy. We introduce the Mixture-of-World Models (MoW), a scalable architecture that integrates three key components: i) modular VAEs for task-adaptive visual compression, ii) a hybrid Transformer-based dynamics model combining task-conditioned experts with a shared backbone, and, iii) a gradient-based task clustering strategy for efficient parameter allocation. On the Atari 100k benchmark, \textbf{a single MoW agent} (trained once over Atari $26$ games) achieves a mean human-normalized score of $\mathbf{110.4}$%, competitive with the $\mathbf{114.2}$% achieved by the recent STORM—an ensemble of $26$ task-specific models—while requiring $50$% fewer parameters. On Meta-World, MoW attains a $\mathbf{74.5}$% average success rate within 300k steps, establishing a new state-of-the-art. These results demonstrate that MoW provides a scalable and parameter-efficient foundation for generalist world models. Our code is available in the supplementary materials.

The Forward-Forward (FF) Algorithm is a recently proposed learning procedure for neural networks that employs two forward passes instead of the traditional forward and backward passes used in backpropagation. However, FF remains largely confined to supervised settings, leaving a gap at domains where learning signals can be yielded more naturally such as RL. In this work, inspired by FF's goodness function using layer activity statistics, we introduce Action-conditioned Root mean squared Q-Functions (ARQ), a novel value estimation method that applies a goodness function and action conditioning for local RL using temporal difference learning. Despite its simplicity and biological grounding, our approach achieves superior performance compared to state-of-the-art local backprop-free RL methods in the MinAtar and the DeepMind Control Suite benchmarks, while also outperforming algorithms trained with backpropagation on most tasks.

Pre-training world models on large, action-free video datasets offers a promising path toward generalist agents, but a fundamental flaw undermines this paradigm. Prevailing methods train models to predict a single, deterministic future, an objective that is ill-posed for inherently stochastic environments where actions are unknown. We contend that a world model should instead learn a structured, probabilistic representation of the future where predictive uncertainty correctly scales with the temporal horizon. To achieve this, we introduce a pre-training framework, **H**orizon-c**A**librated **U**ncertainty **W**orld **M**odel (HAUWM), built on a probabilistic ensemble that predicts frames at randomly sampled future horizons. The core of our method is a Horizon-Calibrated Uncertainty (HCU) loss, which explicitly shapes the latent space by encouraging predictive variance to grow as the model projects further into the future. This approach yields a latent dynamics model that is not only predictive but also equipped with a reliable measure of temporal confidence. When fine-tuned for downstream control, our pre-trained model significantly outperforms state-of-the-art methods across a diverse suite of benchmarks, including MetaWorld, the DeepMind Control Suite, and RoboDesk. These results highlight the critical role of structured uncertainty in robust decision-making.

Large language models (LLMs) are increasingly developed as autonomous agents using reinforcement learning (agentic RL) that reason and act in interactive environments. However, sparse and sometimes unverifiable rewards make it extremely challenging to assign credit when training LLM agents that serve as a policy. Recent work attempts to integrate process supervision into RL but suffers from biased annotation, reward hacking, high-variance from overly fine-grained rewards or failures when state overlap is rare. We therefore introduce implicit step rewards for agentic RL (**iStar**), a general credit-assignment strategy that integrates seamlessly with standard RL algorithms without relying on additional rollouts or explicit step labels. Particularly, we alternatively optimize an implicit process reward model (PRM) with the policy model to generate step rewards for each action via a multi-turn DPO objective. Theoretical analysis shows that this learning objective produces a step-wise reward function learned from trajectory preferences. Then the implicit step rewards are used to compute step-level advantages, which are combined with trajectory (or episode)-level advantages for policy updates, creating a self-reinforcing training loop. We evaluate our method on three challenging agent benchmarks, including WebShop and VisualSokoban, as well as open-ended social interactions with unverifiable rewards in SOTOPIA. Crucially, our method shows superior performance over frontier LLMs and strong RL baselines across domains, achieving state-of-the-art results with higher sample-efficiency and training stability. Further analysis also demonstrates efficient exploration by **iStar** with increased rewards in both step- and episode-level while maintaining fewer steps to achieve task success.

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Large Language Models (LLMs) by optimizing them against factual outcomes. However, this paradigm falters in long-context scenarios, as its reliance on internal parametric knowledge is ill-suited for tasks requiring contextual grounding—the ability to find and reason over externally provided information. We identify a key reason for this failure: a reward based solely on the final answer is too sparse to effectively guide the model for identifying relevant evidence. We formally prove that the outcome-only reward leads to exponentially vanishing gradients for the context grounding process, rendering learning intractable. To overcome this bottleneck, we introduce LongRLVR to augment the sparse answer reward with a dense and verifiable context reward. This auxiliary signal directly incentivizes the model for selecting the correct grounding information, providing a robust learning gradient that solves the underlying optimization challenge. We validate our method on challenging long-context benchmarks using Qwen and LLaMA models. LongRLVR consistently and significantly outperforms the standard RLVR across all models and benchmarks, e.g., boosting a 14B model's scores on RULER-QA from 73.17 to 88.90 and on LongBench v2 from 39.8 to 46.5. Our work demonstrates that explicitly rewarding the grounding process is a critical and effective strategy for unlocking the full reasoning potential of LLMs in long-context applications.