NeurIPS 2026 Papers

Self-play reinforcement learning trains language models on their own generated tasks, co-evolving a proposer and solver without human labels. Recent systems report strong reasoning gains, but collapse and instability are widely observed and poorly understood. The dominant response treats this as a reward-design problem. We argue instead that self-play stability is governed by two distinct levers: a data-level gate that decides which proposer-generated tasks enter the training pool, and the reward signal that updates the policy on tasks already admitted. Through controlled experiments on a Python output-prediction task and a deterministic-DSL twin task that strips pretraining priors, output ambiguity, and executor noise, we find the two levers are asymmetric. A strict gate is sufficient for stability under every reward variant we test, including a self-consistency reward with no access to ground truth; while no reward variant is sufficient once the gate is removed. This asymmetry exposes a counter-intuitive coupling we call the Grounded Proposer Paradox: a proposer with ground-truth access accelerates collapse faster than an ungrounded one when paired with a self-consistency solver, by concentrating training on clean tasks that form the fastest path to a spurious self-consistent attractor. Replacing the binary gate with a continuous strictness parameter $\varepsilon$ further reveals a two-stage phase transition: training-side metrics decouple at low $\varepsilon$, while validation accuracy holds until $\varepsilon$ is much higher. Data-level gating, not reward calibration, is the binding constraint on self-play stability.

Constrained Markov decision processes (CMDPs) extend standard MDPs by incorporating a cost function and enforcing constraints---most commonly by requiring that the expected cumulative cost over the entire trajectory remains below a prescribed threshold. While this global, expectation-based formulation is natural and widely studied, it is insufficient for many safety- and reliability-critical applications, where guarantees are required not only from the initial state but also for every subprocess starting at any intermediate time step. Motivated by this need, we introduce \emph{subprocess-constrained MDPs} (S-CMDPs), in which the expected cumulative cost must satisfy a constraint not only from the initial state, but from every intermediate step onward along the trajectory. We investigate the algorithmic aspects of this model and characterize the limitations of classical solution concepts. In particular, we show that stationary policies can be both suboptimal and computationally intractable to optimize in S-CMDPs. Leveraging a value-set iteration framework, we further demonstrate that history-dependent policies surprisingly overcome both obstacles. Building on this insight, we develop a polynomial-time algorithm that computes the action distribution of a (near-)optimal history-dependent policy for any given history.

Protein optimization remains a longstanding goal in life sciences. Existing machine learning–assisted directed evolution (MLDE) methods primarily rely on sequence-only features, overlooking the critical spatial constraints and co-evolutionary interactions encoded in protein structures. However, directly integrating structural information remains challenging due to the scarcity of reliable mutant structures. To address these issues, we propose **StructEvo**, a novel structure-aware reinforcement learning framework for protein directed evolution. StructEvo employs a delta-structure fusion encoder to approximate mutant structure features via feature differences, enabling dynamic incorporation of spatial knowledge. The vast mutation space is then decomposed into manageable subspaces through a structure-aligned hierarchical action network, while a geometric constraint further stabilizes delta feature learning. Our approach outperforms prior state-of-the-art methods by 9.2% and 16.3% on two challenging optimization benchmarks, and further identifies an experimentally validated epistasis pattern in GFP, highlighting the importance of structural guidance for effective protein directed evolution.

Communication learning is crucial for complex coordination in multi-agent reinforcement learning (MARL), yet existing methods often struggle with a fundamental dilemma between communication overhead and task performance. Typically designed without theoretical guidance, they produce unstructured and redundant communication graphs, limiting both efficiency and scalability. To address this, we propose Structural Entropy Optimized Communication (SEOC), a framework that directly optimizes inter-agent communication topology using the information-theoretic principle of Structural Entropy. SEOC guides the network to self-organize into a low-entropy state with sparse connectivity and distinct communities, enabling efficient and interpretable decentralized coordination. This is achieved through a unified mechanism that jointly enforces structural order and semantic conciseness, reducing topological complexity via a differentiable structural entropy loss while compressing messages with a node information bottleneck. Experiments on benchmarks such as the StarCraft Multi-Agent Challenge (SMACv2) demonstrate that SEOC significantly outperforms state-of-the-art methods.

Tabular prediction has long been dominated by gradient-boosted decision trees and specialized deep tabular models, while large language models (LLMs) remain difficult to make competitive despite their cross-task adaptability and transparent reasoning traces. Existing LLM-for-tabular methods mainly adapt tables into textual inputs, but often fail to capture key structural properties of tabular data. In this work, we propose Permutation Relative Policy Optimization (PRPO), a reinforcement learning post-training method that strengthens LLMs for tabular prediction by injecting column-permutation invariance as a structural prior. By constructing label-preserving column permutations and estimating advantages both within and across them, PRPO converts sparse outcome rewards into denser and more stable optimization signals. Extensive experiments on 139 OpenML datasets show that our 8B model reaches a genuinely competitive regime against strong specialized tabular baselines. It achieves strong fully supervised performance, dominates cross-dataset zero-shot settings, and performs on par with 32-shot strong baselines. Moreover, it substantially outperforms much larger general-purpose and reasoning LLMs, including up to a 53.17% improvement over DeepSeek-R1 (685B). These results show that structural-prior RL post-training is an effective route for making LLMs competitive in tabular prediction.

Machine learning (ML) predictions are increasingly being used to guide decision-making, giving rise to the problem of decision-focused learning (DFL) where predictors are optimized for downstream decision quality rather than accuracy alone. However, existing work assumes a single decision-maker optimizing in isolation. This paper formalizes *strategic* decision-focused learning, where an ML system predicts an exogenous state that some agents observe before playing a game. For example, a park ranger may predict wildlife locations to allocate anti-poaching patrols against strategic poachers. While the exogenous state is unaffected by agent actions, predictions influence agents' strategies and the resulting equilibrium. We find that strategic considerations fundamentally change the learning problem. In particular, we show the prediction accuracy–equilibrium payoff landscape can be non-monotonic---i.e., better predictions can degrade performance. We propose algorithmic approaches to address these challenges and validate them across benchmarks in wildlife conservation and infrastructure protection. Our theory and experiments highlight the importance of accounting for strategic interactions when designing predictors.

Large language models (LLMs) are increasingly used as interactive agents, but optimizing them for long-horizon decision making remains difficult because current methods are largely purely reactive, which weakens both exploration and credit assignment over extended trajectories. In this work, we present Strategic Trajectory Abstraction (StraTA), a simple framework that introduces an explicit trajectory-level strategy into agentic reinforcement learning (RL). StraTA samples a compact strategy from the initial task state, conditions subsequent actions on that strategy, and trains strategy generation and action execution jointly with a hierarchical GRPO-style rollout design, further enhanced by diverse strategy rollout and critical self-judgment. Experiments on ALFWorld, WebShop, and SciWorld show that StraTA consistently improves both sample efficiency and final performance over strong baselines. StraTA reaches success rates of 93.1% on ALFWorld and 84.2% on WebShop, surpassing the strongest baselines by 2.3% and 11.4% respectively. On SciWorld, StraTA attains a 63.5% overall score, outperforming frontier closed-source models by 6.1% and prior RL baselines by 6.5%.

Large reasoning models improve with more test-time computation, but often overthink, producing unnecessarily long chains-of-thought that raise cost without improving accuracy. Prior reinforcement learning approaches typically rely on a single outcome reward with trajectory-level length penalties, which cannot distinguish essential from redundant reasoning steps and therefore yield blunt compression. Although recent work incorporates step-level signals, such as offline pruning, supervised data construction, or verifier-based intermediate rewards, reasoning length is rarely treated as an explicit step-level optimization objective during RL. We propose Step-wise Adaptive Penalization (SWAP), a fine-grained framework that allocates length reduction across steps based on intrinsic contribution. We estimate step importance from the model's on-policy log-probability improvement toward the correct answer, then treat excess length as a penalty mass redistributed to penalize low-importance steps more heavily while preserving high-importance reasoning. We optimize with a unified outcome–process advantage within group-relative policy optimization. Extensive experiments demonstrate that SWAP reduces reasoning length by 56.3% on average while improving accuracy by 4.2% relative to the base model.

Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve the reasoning capabilities of large language models, but its reward is derived only from the correctness of the final answer and provides no supervision over intermediate reasoning steps. Recent rubric-based methods, such as Rubrics as Rewards (RaR), introduce finer-grained supervision by scoring rollouts against a structured set of evaluation criteria. However, the resulting rubric scores are still aggregated into a single scalar that is applied to the entire response, leading to three structural weaknesses, namely loss of the multi-criterion rubric structure, uniform supervision of correct and incorrect reasoning steps, and reward hacking in the trained model through unbounded self-correction. On a sample of 1{,}000 problems, we find that 18.2\% of steps within answer-correct responses are themselves wrong yet positively rewarded, while 49.9\% of steps within answer-incorrect responses are in fact correct yet penalized. Therefore, we introduce \textbf{Step-wise Rubrics as Rewards (SRaR)}, an RLVR framework that (i) uses an LLM judge to attribute each rubric item to a specific reasoning step, (ii) normalizes the per-step rubric scores across rollouts so that only steps whose quality varies produce a learning signal, and (iii) combines the resulting per-step reward with the standard outcome reward through a decoupled advantage estimator that keeps the outcome-driven baseline stable. To support training, we further build a 16K-problem rubric dataset by contrastively distilling rubric items from correct and flawed reasoning paths sampled from a strong model and verified against the ground-truth answer. Across six mathematical reasoning benchmarks spanning multiple difficulty levels, SRaR improves the average accuracy over RaR by \textbf{3.57} points on Qwen3-8B-Non-Thinking and by \textbf{2.75} points on Qwen3-32B-Non-Thinking, raises the Faithful Reasoning Rate on AIME~2025 from 34.5\% to 46.7\% (every correct-answer trajectory of SRaR uses entirely correct reasoning steps), and reduces the rate of self-correction looping from 48.1\% to 26.5\%.

Verifiable reward training has improved mathematical and coding reasoning, but these domains capture only part of step-by-step decision making. Many real-world tasks require finding a high-value feasible plan among many valid alternatives. We introduce $\text{OPT}\star$, a scalable family of optimization-style tasks for training and evaluating LLM step-by-step optimization-like reasoning along a complexity axis: each task provides a feasibility checker and evaluator, while a complexity parameter expands the search space without requiring new human labels. This motivates studying these tasks in two regimes: (i) solver-guided online policy optimization, which uses a solver as a value oracle for partial states and applies rank-based reward shaping to reinforce better next steps, and (ii) search-based offline RL when such solvers are unavailable. Theoretically, we relate success in large search spaces to the information a reasoner extracts per unit of search budget. Empirically, we ablate the ingredients that make search efficient on $\text{OPT}\star$ and show that training on $\text{OPT}\star$ improves step-by-step optimization-like reasoning.

Steering methods influence Large Language Model behavior by identifying semantic directions in hidden representations, and are typically realized through inference-time activation interventions that apply a fixed, global modification to the model's internal states. While effective, such interventions often induce unfavorable attribute--utility trade-offs under strong control, as they ignore the fact that many behaviors are governed by a small and heterogeneous subset of model components. To alleviate the trade-offs, we propose Steer2Edit, a theoretically grounded, training-free framework that transforms steering vectors from inference-time control signals into diagnostic signals for component-level rank-1 weight editing. Instead of uniformly injecting a steering direction during generation, Steer2Edit selectively redistributes behavioral influence across individual attention heads and MLP neurons, yielding interpretable edits that preserve the standard forward pass and remain compatible with optimized parallel inference. Across multiple tasks including safety alignment, truthfulness promotion, and reasoning efficiency, Steer2Edit consistently achieves a superior attribute--utility trade-offs: at matched downstream performance, it outperforms the strongest steering baseline by 17.2% in safety, 9.8% in truthfulness, and 12.2% in reasoning length reduction. Overall, Steer2Edit provides a principled bridge between representation steering and weight editing by translating steering signals into interpretable, training-free parameter updates.

Bellman residual minimization (BRM) provides a scalable, gradient-based approach to offline reinforcement learning in large state spaces. While globally convergent gradient-based soft (i.e., entropy-regularized) BRM methods for neural networks have recently been established, their statistical complexity under stochastic gradient descent remains largely unknown. In this paper, we address this theoretical gap for soft BRM. Through a novel Lyapunov-based analysis, we establish an $\mathcal{O}(1 / n)$ average argument stability bound, which translates directly into a $\mathcal{O}(1 / n)$ statistical complexity for the soft BRM objective.

Large language models deployed as personalized assistants must reason over long, evolving interaction histories. However, in long-term dialogue reasoning, relevant evidence is scattered across sessions, preferences may be revised over time, and standard long-context training fails to address these challenges under data scarcity and prohibitive computational costs. We propose StateTree, a data-driven RL pseudo-task that constructs a challenging auxiliary task from scarce dialogues with verifiable ground truth. StateTree augments multi-session dialogues with a tree-structured path-tracing task: key-value records are embedded across sessions to form a binary tree. Solving the task requires the model to traverse from root to leaf by retrieving records across sessions and comparing timestamps to resolve branches, then recover the hidden target question among distractor leaves. We apply curriculum RL training progressively increasing tree depth and introduce a compositional variant whose edges carry step-level reasoning fragments, training the model to compose partial cues into coherent queries. Trained on 10K-token contexts, StateTree generalizes to 128K tokens without full-length RL costs and exhibits capabilities including cross-session retrieval, temporal reasoning, knowledge update, and compositional multi-hop reasoning. StateTree outperforms both SFT and RL-based baselines while preserving short-context general reasoning. StateTree-7B achieves gains up to +23.60% on LongMemEval (128k), and StateTree-14B reaches 59.00% accuracy on LongMemEval, surpassing QwenLong-L1-32B (45.20%).

Immunohistochemical (IHC) staining encodes molecular protein expression critical for clinical diagnosis, yet its chemical procedures are costly and time-consuming. Virtual staining offers a compelling alternative by digitally synthesizing IHC images from hematoxylin-eosin (H\&E) stained slides. Despite remarkable progress in diffusion- and flow-matching-based virtual staining, supervised fine-tuning (SFT) remains fundamentally limited in pathological fidelity due to its coarse, spatially averaged supervision signal that starves sparse DAB-positive regions of effective gradients. Reinforcement learning (RL) can exploit inter-rollout variance for outcome-level pathological supervision, but naively applying DAB-based rewards triggers severe reward hacking that collapses the image distribution and degrades perceptual quality. This paper presents StainNFT, a flow matching RL post-training framework built on a curriculum reward strategy that gates fine-grained optical density supervision on per-sample DAB mask IoU. To enhance fine-grained pathological fidelity, we introduce expression-aware reweighting, multi-scale block supervision, and a closed-loop implicit pathological semantic reward derived from pathology foundation model priors. Extensive experiments on seven benchmarks demonstrate that StainNFT consistently outperforms existing methods in both perceptual quality and pathological fidelity for H\&E-to-IHC virtual staining. Ablation studies confirm the effectiveness of each proposed component. Our code and trained models will be released.

Generative recommendation (GR), which directly generates sequential semantic IDs for items, has recently attracted increasing attention in recommender systems. Given that core recommendation metrics are inherently evaluated at the sequence level, reinforcement learning (RL) serves as a natural optimization framework. However, adapting recent group-based RL algorithms to GR presents notable challenges: token-level methods such as GRPO exhibit much more severe training instability, while sequence-level methods such as GSPO enforce uniform token weighting and thus ignore the heterogeneous roles of tokens. To address this dilemma, we propose Stable and Granular Policy Optimization (SGPO), a group-based RL algorithm tailored for GR. SGPO maintains a sequence-level ratio to promote optimization stability while injecting fine-grained token-wise advantages. By incorporating token log-probabilities and reward signs, SGPO explicitly upweights high-confidence tokens in positive trajectories to reinforce successful retrieval, while strictly penalizing them in negative trajectories to encourage exploration, and uses an adaptive scaling factor to bound the variance of reshaped advantages. Experiments on two public benchmarks and a billion-scale industrial dataset show that SGPO consistently outperforms six strong RL baselines and, when deployed on a commercial platform, achieves a 2.36% lift in click-through rate (CTR) and a 1.95% increase in advertising revenue in online A/B testing.

While reinforcement learning (RL) has been central to the recent success of large language models (LLMs), RL optimization is notoriously unstable, especially when compared to supervised fine-tuning (SFT). In this work, we investigate the stability gap between SFT and RL from a gradient-based perspective, and show that the convexity of the SFT loss with respect to model logits plays a key role in enabling stable training. Our theoretical analysis demonstrates that this property induces favorable gradient directionality during optimization. In contrast, Proximal Policy Optimization (PPO), a widely adopted policy gradient algorithm utilizing a clipped surrogate objective, lacks this stabilizing property. Motivated by this observation, we propose Logits Convex Optimization (LCO), a simple yet effective policy optimization framework that aligns the learned policy with an optimal target derived from the original RL objective, thereby emulating the stabilizing effects of logits-level convexity. Extensive experiments across multiple model families show that our LCO framework consistently improves training stability and outperforms conventional RL methods on a broad range of benchmarks.

Reinforcement Learning with Verifiable Rewards (RLVR) has become an effective post-training method for improving the reasoning abilities of Large Language Models (LLMs). However, existing methods mainly apply uniform optimization constraints across all tokens, ignoring their heterogeneous roles. Prior work shows that high-entropy tokens are closely tied to reasoning, while low-entropy tokens primarily encode factual knowledge, and recent approaches attempt to exploit this distinction by isolating token updates via masking or asynchronous training. We argue that such isolation breaks the sequential dependency structure of autoregressive generation, leading to suboptimal learning. To address this, we propose \textbf{Archer}, an entropy-aware RLVR framework with \textbf{dual-token constraints} that preserves joint optimization while modulating update strength across token types. Our method introduces response-level entropy normalization for stable token classification and applies differentiated clipping ranges and KL regularization to encourage exploration on reasoning tokens while preserving knowledge tokens. Experiments on mathematical reasoning and code generation benchmarks show that Archer consistently outperforms strong baselines across multiple model scales, improving both \textit{pass@1} and \textit{pass@K} performance. These results highlight the importance of respecting sequence-level dependencies when designing fine-grained RL optimization strategies for LLMs.

We introduce Split-RL, a reinforcement learning (RL) algorithm based on decision trees. It uses guidance labels to explicitly partition the state-space and isolate localized conflicting signals. RL policies are often trained by mixing contradictory feedback that is localized in space and time. For example, a robot may prioritize speed in open areas versus precision in tight gaps. These regions are often easily detectable from sensor data or a human operator. Standard neural networks (NNs) lack the structural mechanism to isolate these signals. While Gradient Boosting Trees (GBT) provides the inductive bias to explicitly partition the state-space, standard tree-fitting procedures fit only aggregated gradients, merging contradictory updates before they reach the leaves. Split-RL leverages this bias to integrate guidance labels into tree construction, routing these updates into disjoint regions of the state-space to prevent signal interference during training. Such labels can arise from simple heuristics, sensor-derived events, or existing rule-based systems, allowing Split-RL to incorporate rule-derived structure while retaining a learned policy. We theoretically show how early gradient aggregation loses objective-specific information and characterize when the Split-RL score isolates conflicting updates. Finally, we evaluate Split-RL on constrained tasks and offline imitation learning from mixed-quality datasets. Across constrained domains with spatial, rare-event, and temporal conflicts, Split-RL consistently outperforms existing methods by achieving low-cost, feasible solutions while maintaining competitive rewards. In offline settings, Split-RL localizes the tradeoff between imitation and policy improvement, outperforming other methods and achieving near-expert performance across varying dataset sizes.

This work investigates multi-objective imitation learning: the problem of recovering policies that lie on the Pareto front given demonstrations from multiple Pareto-optimal experts in a Multi-Objective Markov Decision Process (MOMDP). Standard imitation approaches are ill-equipped for this regime, as naively aggregating conflicting expert trajectories can result in dominated policies. To address this, we introduce Multi-Output Augmented Behavioral Cloning (MA-BC), an algorithm that systematically partitions divergent expert data while pooling state-action pairs where no behavior conflict is observed. Theoretically, we prove that MA-BC converges to Pareto-optimal policies at a faster statistical rate than any learner that considers each expert dataset independently. Furthermore, we establish a novel lower bound for multi-objective imitation learning, demonstrating that MA-BC is minimax optimal. Finally, we empirically validate our algorithm across diverse discrete environments and, guided by our theoretical insights, extend and evaluate MA-BC on a continuous Linear Quadratic Regulator (LQR) control task.