NeurIPS 2026 论文列表

Process reward models enable fine-grained, step-level evaluation of LLMs, yet building them for agentic settings remains prohibitively difficult: long-horizon interactions, irreversible actions, and stochastic environment feedback make both human annotation and Monte Carlo estimation infeasible at scale. In this work, we show that reinforcement learning (RL) post-training already provides the ingredients for effective step-level scoring, eliminating the need for dedicated reward model training altogether. Concretely, we derive an implicit advantage under a general stochastic Markov Decision Process, which we term progress advantage---log-probability ratio between the RL-trained policy and its reference policy exactly recovers the optimal advantage function. This signal is annotation-free, domain-agnostic, and available as a byproduct of the standard RL post-training pipeline. We validate the effectiveness of the progress advantage across three different applications: test-time scaling, uncertainty quantification, and failure attribution on five benchmarks and four model families. Across all settings, progress advantage consistently outperforms confidence-based baselines and, despite requiring no task-specific training, surpasses dedicated trained reward models. We complement these results with deeper analyses on characteristics of progress advantage, offering practical guidance for adoption in real-world agentic systems.

In emotional support conversations, user needs are dynamic, shifting across turns between emotional soothing, cognitive guidance, and their combinations. Existing methods often optimize a single user-state objective, such as emotional relief or cognitive growth, making it difficult to capture such need shifts. In this paper, we formulate emotional support dialogue as a need-aware multi-objective optimization problem, where agents learn to generate responses that balance multiple support objectives according to users' evolving needs. We propose \textbf{MindFlow}, a psychologically grounded user simulator that tracks users' emotional and cognitive needs at each turn, derives state-dependent preference vectors from personalized arousal curves, and models mental-state transitions with a dual-process system. Building on MindFlow, we introduce \textbf{NeMo}, a need-aware multi-objective reinforcement learning framework that estimates local multi-turn effects through forward simulation and dynamically aggregates emotional and cognitive rewards based on current support preferences. By converting delayed, multi-dimensional feedback into step-level optimization signals, NeMo improves credit assignment and aligns policy learning with long-term user-state improvement. Experiments show that NeMo consistently improves user-side state metrics and response-level supportive quality, with robust generalization under both Sim2Sim and Sim2Real evaluations. Code and data are available at https://anonymous.4open.science/r/NeMo-6EB9.

We study fixed-confidence best-arm identification under strict 1-bit feedback constraints. At each round, the learner selects an arm and a query set, and receives only a single bit indicating whether the sampled reward belongs to that set. We consider a distribution-free finite-variance setting with arm-wise localization, where direct empirical mean estimation is no longer available and clipping becomes unavoidable. We first formulate a time-uniform 1-bit mean-estimation primitive based on randomized threshold queries and a clipped tail-integral identity. We then embed this primitive into candidate-challenger best-arm identification algorithms. A fixed-clipping algorithm gives a simple anytime $(\epsilon,\delta)$-PAC guarantee, while a phased adaptive-clipping algorithm matches the clipping level to the current resolution and yields a gap-adaptive sample complexity. We also prove a two-arm information-theoretic lower bound showing that the logarithmic penalty caused by finite-variance 1-bit feedback is intrinsic. Consequently, the phased algorithm is optimal in the leading dependence up to lower-order $\log\log$ factors.

Vision-Language-Action (VLA) models are formulated to ground instructions in visual context and generate action sequences for robotic manipulation. Despite recent progress, VLA models still face challenges in learning related and reusable primitives, reducing reliance on large-scale data and complex architectures, and enabling exploration beyond demonstrations. To address these challenges, we propose a novel \textbf{N}euro-\textbf{S}ymbolic \textbf{V}ision-\textbf{L}anguage-\textbf{A}ction (NS-VLA) framework via online reinforcement learning (RL). It introduces a symbolic encoder to embedding vision and language features and extract structured primitives, utilizes a symbolic solver for data-efficient action sequencing, and leverages online RL to optimize generation via expansive exploration. Experiments on robotic manipulation benchmarks demonstrate that NS-VLA outperforms previous methods in both one-shot training and data-perturbed settings, while simultaneously exhibiting superior zero-shot generalizability, high data efficiency, and an expanded exploration space. Our code is available at https://anonymous.4open.science/r/NS-VLA/.

Augmenting model-free reinforcement learning (RL) with representations learned through observation dynamics prediction (observation-predictive RL) can improve sample efficiency and performance, with minor modifications and limited additional computation. However, this approach still struggles in challenging tasks with low-dimensional observations. In this paper, we identify a key factor behind this problem: unbalanced reconstruction losses across observation dimensions, where dimensions with larger value ranges dominate the loss. This encourages the agent to neglect dimensions with relatively small ranges, leading to degraded performance. To address this issue, we propose a novel normalization method tailored to online RL, which normalizes low-dimensional observations and balances the resulting losses and gradients. Beyond balancing reconstruction losses, observation normalization enables dynamics prediction to be performed in a normalized observation space, thereby providing a unified treatment of low- and high-dimensional inputs (e.g., physical states and images). Building on this idea, we further introduce Normalized Observation Space Dynamics-Augmented Q-learning (NASDAQ), a framework for observation-predictive RL applicable across diverse domains. NASDAQ learns state-action representations by coupling value learning with two auxiliary tasks: short-term value prediction and next *normalized* observation prediction. Extensive experiments demonstrate that NASDAQ achieves competitive or superior performance compared with state-of-the-art model-based and self-predictive RL methods, while requiring significantly less training wall-time.

Mathematical optimization drives decisions in supply chains, logistics, energy, and scheduling, but translating natural-language problems into solver-executable formulations remains a bottleneck. While frontier models show promise for this task, many deployments require small language models (SLMs) that run locally under privacy, latency, and cost constraints. Existing SLMs are limited by noisy data, ambiguous benchmarks, and one-shot inference that ignores the iterative, feedback-driven nature of optimization modeling. We introduce \textsc{OptiMind}, a framework for improving SLMs through formulation-centric data cleaning, class-conditional priors, and multi-turn reinforcement learning with solver feedback. \textsc{OptiMind} repairs ambiguous problems, regenerates solution trajectories, audits benchmark labels, constructs accepted objective-value sets, and distills class-level hints from recurring formulation errors. It then trains models to formulate, execute, and revise GurobiPy programs using solver-verified correctness as reward. On expert-cleaned IndustryOR, Mamo-Complex, and OptMATH benchmarks, a 20B \textsc{OptiMind-RL} model outperforms strong open-source baselines and reaches frontier-model performance under the same protocol. Our results suggest that optimization formulation is better treated as an interactive solver-grounded decision process than as one-shot code synthesis. We release our inference framework, cleaned benchmarks, and sample error analyses at \url{https://anonymous.4open.science/r/OptiMind-NeurIPS-1DB020}, with full data and model to follow.

Reinforcement learning with verifiable rewards (RLVR) plays a pivotal role in improving the reasoning ability of large language models. However, widely used PPO surrogate objectives are fundamentally *local*, as they rely on a local approximation of the exact policy gradient objective. While this approximation improves stability by reducing the variance induced by importance sampling, it also introduces structural bias into the surrogate objective, which must be controlled through trust region mechanisms. In this work, we introduce the $N$-step forward trace, which augments the PPO surrogate objective using the cumulative likelihood ratio of the next $N-1$ tokens. Building on this idea, we propose $N$-Step Forward-Trace Policy Optimization (NFPO), a practical RLVR algorithm that integrates the $N$-step forward trace into the masked policy gradient framework. NFPO provides a continuous bridge between the PPO surrogate objective and the exact policy gradient objective, offering a principled mechanism for controlling the bias–variance trade-off. Our theoretical analysis shows that, with an appropriate choice of $N$, the proposed objective yields a tighter policy-improvement bound than the standard PPO surrogate. Experiments on comprehensive reasoning benchmarks demonstrate that NFPO consistently improves performance, supporting our theoretical findings.

Long-horizon planning often requires composing familiar local transitions while remaining grounded in a specified goal. We study this setting from single-stage data: on a controlled no-skip graph benchmark, training provides only adjacent-stage demonstrations while evaluation requires composing them on unseen long-horizon start–goal pairs; an event-chain diagnostic decomposes success into legality, stage-bridging, and goal-conditioned termination. Across a minimal Transformer, Qwen2.5-3B, and an external validation on MuSiQue multi-hop QA, supervised fine-tuning (SFT) learns local transitions but degrades with horizon, with failures localized to stage-bridging. Reinforcement learning (RL) narrows this composition gap, especially at longer horizons, but unanchored RL drifts—producing valid stage progress while ignoring the instructed target. Stable gains require two complementary forms of anchoring: instruction-style prompting strengthens prompt-side goal-conditioning, especially at shorter horizons, whereas KL regularization to the SFT reference policy prevents RL-induced drift across horizons; reward shaping mainly provides denser credit. Finally, we identify a Decomposition Paradox: models trained only on long compositions can reach intermediate targets yet fail the corresponding prefix subtasks because stopping at instructed intermediates is never supervised.

Multi-objective reinforcement learning (MORL) is a fundamental problem for modern AI, with applications spanning robotics, LLMs, and reasoning. A central challenge is to learn policies that balance competing objectives, such as speed versus energy in robotics, or helpfulness versus honesty in alignment. Optimal policies can vary sharply with the preferred objective weights, and the severity of this variation depends crucially on the degree of misalignment among objectives, making per-preference learning sample-inefficient. To formalize this, we introduce an *objective-conflict coefficient* and use it to derive near-optimal regret bounds for online MORL that quantify the impact of misalignment on learning efficiency. Motivated by this analysis, we propose *routed ensemble policy learning*, an algorithm that maintains a small set of basis policies and dynamically routes trajectories to the policy with the highest potential reward. We further design a scalable, kernel-based attribution algorithm that estimates how feedback signals from one set of trajectories influence policy learning on others. Unlike existing attribution methods, our approach captures nonlinear interactions, such as synergy and antagonism between trajectories, to explain objective conflict. We extensively evaluate our approach across diverse environments and benchmarks: routed ensembles improve average performance by 4.3\% over MORL baselines on several environments from the Meta-World and MO-Gymnasium benchmarks, and kernel-based attribution improves correlation with leave-one-out analysis over existing baselines by 50\%. When applied to downstream task selection, kernel-based attribution yields a 27\% improvement in reward over baseline attribution methods.

Direct preference optimization (DPO) is a simple and effective alignment strategy for large language models (LLMs) based on pairwise preferences between two candidates. In recommender systems, however, user feedback is rarely pairwise. For a given context, e.g., a user, a session, or a conversation, we typically observe set-wise preferences with multiple positive items, where every positive item should outrank every unobserved or explicitly negative item, with no prescribed order among the positives or the negatives themselves. A natural generalization is to use the Plackett–Luce (PL) reward model, which extends the Bradley–Terry (BT) reward model underlying vanilla DPO from pairwise preferences to full rankings of candidates. However, adapting the PL model to set-wise preferences requires marginalizing over all consistent permutations of the positives and negatives respectively, which is intractable. To address this fundamental challenge, we propose Mult-DPO, a novel DPO objective with a tractable multinomial reward model over set-wise preferences. We show that, like the BT and PL reward models, the multinomial reward admits a closed-form DPO-style objective, enabling direct alignment of LLMs without reinforcement learning (RL). In addition, we prove that the multinomial DPO loss is a tractable upper bound on the exact marginalized PL DPO loss when optimizing against the set-wise preference data. We further characterize the tightness of this bound in terms of relative total weight (i.e., exponentiated rewards) of positives versus negatives, which provides insights into tightening the bounds with more or harder negatives. Finally, we extend the framework to a group-wise setting that accommodates multiple preference levels. Code and datasets are available at https://anonymous.4open.science/r/mult-dpo-41C7.

Modelling opinion dynamics typically relies on hand-crafted local interaction rules to study emergent macroscopic phenomena such as consensus and polarisation. In contrast, multi-agent reinforcement learning (MARL) enables agents to learn such behaviours directly by optimising simple rewards. To explore the potential of MARL for opinion dynamics, we introduce a GPU-accelerated *consensus and truth-finding game* that scales to populations of up to 1000 agents, comparable to many real-world social sub-networks. To prevent unrealistic conventions, we extend other-play to general-sum social interactions. We next validate our model on a subset of the Bluesky network by recovering agent importance structures from graph topology alone via a learned attention layer, finding that *highly conforming* populations most closely match human data. In large social media networks such high levels of conforming significantly reduce collective accuracy and promote dishonest agents that lie to fit in. By contrast, small, dynamic hunter-gatherer networks are less affected; here, conformity can even improve collective agreement. This suggests a mismatch between evolved human conformity heuristics and modern social media environments as a potential contributor to misinformation.

Variational quantum algorithms offer a promising route to practical quantum advantage on near-term hardware, yet training quantum neural networks (QNNs) remains hampered by two compounding difficulties: the exponential vanishing of gradient variance known as the barren plateau, and the $\mathcal{O}(L \cdot 2^n)$ time–memory cost of differentiating through an $n$-qubit, $L$-layer circuit. We propose RLQ-Grad, a reinforcement-learning-based optimizer in which a classical policy $\pi_\phi$ (a spectrally-normalized PPO agent) learns to propose parameter updates directly, conditioned on the QNN's current parameters, loss, accuracy, and previous update. Because the surrogate gradient is emitted by a classical network rather than obtained by differentiating through the unitary $U(\theta)$, its variance is not constrained by the barren plateau concentration bound, and its per-step cost scales independent of the Hilbert-space dimension. We prove these two properties formally and verify them empirically on a hardware-efficient ansatz across four supervised benchmarks. RLQ-Grad preserves a near-flat gradient-variance curve where backpropagation, parameter-shift, and adjoint differentiation decay by 1–2 orders of magnitude, attains $128\times$–$1841\times$ wall-clock speedups and constant $\sim0.1$ MB memory at $n=12$ qubits, and improves top-1 validation accuracy by up to $+10\%$ over the strongest gradient-based baseline at every circuit scale tested.

We propose and analyze a model-based bootstrap for transition kernels in finite controlled Markov chains (CMCs) with possibly nonstationary or history-dependent control policies, a setting that arises naturally in offline reinforcement learning (RL) when the behavior policy generating the data is unknown. We establish distributional consistency of the bootstrap transition estimator in both a single long-chain regime and the episodic offline RL regime. The key technical tools are a novel bootstrap law of large numbers (LLN) for the visitation counts and a novel use of the martingale central limit theorem (CLT) for the bootstrap transition increments. We extend bootstrap distributional consistency to the downstream targets of offline policy evaluation (OPE) and optimal policy recovery (OPR) via the delta method by verifying Hadamard differentiability of the Bellman operators, yielding asymptotically valid confidence intervals for value and $Q$-functions. Experiments on the RiverSwim problem show that the proposed bootstrap confidence intervals (CIs), especially the percentile CIs, outperform the episodic bootstrap and plug-in CLT CIs, and are often close to nominal ($50\%$, $90\%$, $95\%$) coverage, while the baselines are poorly calibrated at small sample sizes and short episode lengths.

Language models fine-tuned with reinforcement learning frequently learn to exploit their reward source rather than solve the underlying task. Existing fixes are typically reactive: detect a specific exploit, then patch the reward. We instead ask whether reward hacking can be prevented by constraining how the model represents the task. We introduce prompt KL regularization, a single auxiliary loss that adds a KL penalty between the actor's and reference model's per-token distributions on the prompt only, leaving the response distribution free. Across three reward hacking tasks of different sizes, prompt KL regularization maintains low hack rates (under 3\%) when standard training yields up to 100\% hack rates. Additionally, it matches or improves ground-truth accuracy, with no changes to the reward or environment. We aim to understand why prompt KL regularization is effective. Two independent mechanistic studies suggest that changes in prompt representations are key to models learning reward hacking behavior. First, we show that prompt-space activation steering Pareto-dominates response-space steering on hack rate and accuracy. Then, we demonstrate that swapping prompt activations between models can transfer and reduce reward hacking. Together, these results suggest that the way models internally represent the task is central to reward hacking, and that this representation can be directly constrained to prevent it.

Reinforcement learning agents trained to maximize their own reward in repeated interactions can converge to supra-competitive outcomes resembling explicit collusion, without communication or shared design. Existing mitigation approaches are largely tied to specific economic settings, like two-sided platforms and auctions, leaving open how to design interventions for general repeated games. We address this gap by formalizing the connection between empirical observations from prior work on Q-learning collusion and classical theory of Simple Penal Codes (SPCs). We show any non-trivial SPC induces a quantifiable conditional dependence in agents' policies, detectable via the total variation distance between an agent's action distributions across cooperation and defection histories. Building on this connection, we propose CURB (Collusion Unwinding via Reward shaping and Belief injection), a reward-shaping framework that penalizes this Total Variation (TV) distance signal during Q-learning and is guaranteed to convert any SPC fixed point of the dynamics into a trivial one, thus precluding collusive equilibria sustained by punishment threats. Empirically, CURB substantially reduces collusion by Q-learning agents in both Bertrand and Cournot Competition Repeated Games. We further demonstrate that CURB extends to deep Q-network agents in Bertrand competition, suggesting the mechanism generalizes beyond tabular Q-learning.

Random Network Distillation (RND) is a scalable novelty signal for reinforcement learning, but it can overgeneralize: the predictor extrapolates the random target off the data manifold, causing novelty scores to collapse on unfamiliar inputs. This paper analyzes this failure through a kernel-theoretic lens. Under a GP target model and an NTK-regime predictor, we show that the expected RND energy is a cross-kernel residual governed jointly by the target covariance kernel and the predictor interpolation kernel. This identifies overgeneralization as spectral under-excitation of modes where predictor residuals would otherwise survive. We then explicitly formulate spectral target design as a principle for shaping RND residual geometry; one concrete instance is replacing to replace the implicit random-network target with a bandwidth-controlled Random Fourier Feature (RFF) target. Numerical experiments across offline D4RL and online Atari benchmarks show this target-design approach improves novelty discrimination and downstream RL performance.

Large reasoning models (LRMs) improve language model capabilities by generating explicit thinking traces before final answers. In factuality-oriented question answering (QA), such thinking often improves overall performance by helping the model recover relevant knowledge and refine its answers. However, we find that this benefit is not uniform at the instance level: explicit thinking can also overturn correct non-thinking answers and lead to factual drift. We refer to this failure mode as \emph{thinking-induced hallucination}. To explain this phenomenon, we formulate explicit thinking in factuality QA as a thinking residual over the model's direct-answer tendency, which can either recover missing knowledge or introduce unsupported associations. Based on this formulation, we propose MARGO, \underline{\textit{M}}ixed-Mode \underline{\textit{A}}dvantage \underline{\textit{R}}egularization for \underline{\textit{G}}rounded \underline{\textit{O}}ptimization, a reinforcement learning framework that uses non-thinking rollouts as same-model references in advantage estimation. By constructing mixed-mode rollout groups with both thinking and non-thinking trajectories, MARGO evaluates whether explicit thinking adds factual value beyond direct answering, thereby suppressing hallucination-prone thinking while preserving beneficial thinking behaviors. Experiments across multiple factuality-oriented QA benchmarks demonstrate that MARGO improves factual reliability over strong baselines, while evaluations on mathematical benchmarks show that it preserves general reasoning ability.

Model-based reinforcement learning (MBRL) promises high sample efficiency but is often crippled by compounding model errors: small prediction inaccuracies accumulate rapidly over multi-step rollouts, leading to biased policy optimization. We propose FTA, a trajectory-level generative framework using conditional flow matching (CFM) instead of step-wise dynamics. To avoid curved, energy-inefficient paths, we introduce an optimal transport regularizer grounded in the Benamou–Brenier formula. This regularizer minimizes kinetic energy while enforcing exact state displacement, yielding straight, low-energy trajectories with provable linear error propagation. Since a static generator fails in online RL due to (i) shifting data distributions and (ii) mismatch between training (historical) and desired (high-return) conditionals, we add periodic retraining from the replay buffer and a value-guided ODE sampler that biases generation toward high returns via the current Q-function. Theoretically, the regularizer drives the flow toward the Wasserstein-2 optimal transport map, and value-guided sampling reduces asymptotic Q-estimate bias. On MuJoCo benchmarks, FTA consistently outperforms strong MBRL and model-free baselines in sample efficiency and final performance, and generalizes across tasks. By mitigating compounding errors at the trajectory level and adapting to online shifts, FTA offers a robust, data-efficient alternative to traditional dynamics models.

In cooperative multi-agent reinforcement learning under partial observability, agents typically differ in both their local observations and in how much value those observations let them recover. Most existing methods treat agents symmetrically when shaping credit and exploration, ignoring this asymmetric information disadvantage. We introduce the per-agent value gap, a state- and agent-specific quantity that measures the value loss attributable to local information limits, and show that it serves as a single, measurable signal that unifies credit assignment and exploration: down-weighting policy updates where an agent is at a strong disadvantage, and rewarding behaviors that reduce the disadvantage along the trajectory. We instantiate this idea as \method{}, which estimates the gap online with a per-agent local value head and uses it to drive both advantage shaping and an information-seeking intrinsic reward. We further show that the resulting shaping reduces the variance of the policy-gradient estimate at high-disadvantage states under a mild correlation assumption, and that the intrinsic bonus vanishes in expectation under any stationary optimal policy, so the augmented objective remains consistent with the original cooperative objective. On sixteen tasks across five cooperative benchmarks, \method{} ranks first or statistically tied for first on fifteen of sixteen tasks (strictly first on twelve), with the largest gains on settings combining sparse rewards and asymmetric observability (e.g., a $24\%$ mean relative gain over the strongest baseline averaged across the three SMAC v2 tasks).