NeurIPS 2026 论文列表

Building general-purpose role-playing agents that faithfully portray any character from a natural-language profile remains challenging. The dominant paradigm---supervised fine-tuning---encourages behavioral mimicry without deep, human-like internal thought processes, resulting in poor out-of-distribution generalization. Therefore, we propose Psy-CoT, a psychology-grounded chain-of-thought framework that decomposes pre-response reasoning into three role-specific steps---Interaction Perception, Psychological Empathy, and Logical Construction---so that the model thinks dynamically from the profile rather than merely mimicking surface patterns. While structured reasoning provides a foundation, it alone is insufficient; reinforcement learning is essential to further align the model with character fidelity. However, we observe that under LLM-based reward models, both generic phrases that hack the reward model and genuinely role-specific phrases receive identical gradient signals---this hacking accumulates over training, misleading the model into treating both as equally optimal choices. To address this, we propose Role-Aware Policy Optimization (RAPO), which uses profile--token mutual information to weight gradients asymmetrically---amplifying role-specific tokens under positive advantage while attenuating them under negative advantage. Experiments on CoSER, CharacterBench, and CharacterEval demonstrate that Psy-CoT outperforms existing role-playing CoT methods, and RAPO consistently surpasses GRPO across multiple model scales.

We study regret minimization for time-homogeneous episodic tabular reinforcement learning with $S$ states and $A$ actions under non-interactive $\epsilon$-local differential privacy (LDP). In each episode $k \in [K]$, the decision-maker sends a policy to a user, the user executes the policy for $H$ steps, and sends an $\epsilon$-locally differentially private message back to the decision-maker. When $H=1$, the problem reduces to an LDP contextual bandit with $S$ contexts, $A$ arms and horizon $K$. We prove that the minimax regret is $\min\bigg(K, \frac{ \sqrt{\max( S, \exp(\epsilon) + 1) S A K (\exp(\epsilon) + 1)}}{\exp(\epsilon) - 1}\bigg)$, up to constants and a $\sqrt{\log(A)}$ factor. The upper bound is achieved by an adaptation of EXP3 and a unary encoding mechanism. When $H > 1$, and for small $\epsilon$, we show a lower bound in $\Omega(HS\frac{\sqrt{A K}}{\epsilon})$ and an upper bound in $\tilde{O}(H^2 S^2 \frac{\sqrt{ AK}}{\epsilon})$. The upper bound is achieved by an algorithm based on UCBVI and a unary encoding mechanism. For small $\epsilon$, these bounds improve the previously known lower bound by a factor $\sqrt{S}$ and the previously known upper bound by a factor of $H \sqrt{A}$.

This paper studies the regret analysis for parallel Gaussian process (GP) bandit optimization. The known regret upper bounds for the widely used GP batched upper confidence bound and GP batched Thompson sampling (GP-BTS) suffer from a multiplicative factor with respect to the batch size $Q$. To avoid this degradation, existing analyses require a polynomial number of uncertainty sampling (US) for $Q$ at the beginning of optimization. However, this initial US phase is often ineffective in practice. This paper shows that the regret upper bound without the multiplicative factor on $Q$ can be achieved without the initial US phase, using GP-BTS as an example. Furthermore, we show much better regret upper bounds in the noiseless setting than in the noisy setting, as in the sequential GP bandit setting.

We propose a novel definition of model exploitation in reinforcement learning. Informally, a world model is exploitable if it implies that one policy should be strictly preferred over another while the environment’s true transition model implies the reverse. We analogize our definition with a prior characterization of reward hacking but show that the associated proof of inevitability does not transfer to exploitation. To overcome this obstruction, we develop a general theory of reward hacking and model exploitation that proves that exploitation is essentially unavoidable on large policy sets and yields the corresponding claim for hacking as a special case. Unfortunately, we also find that the conditions that guarantee unhackability in finite policy sets have no counterparts that preclude exploitation. Consequently, we introduce a relaxed notion of exploitation and derive a safe horizon within which it can be avoided. Taken together, our results establish a formal bridge between reward hacking and model exploitation and elucidate the limits of safe planning in world models.

In-context reinforcement learning (ICRL) is often described as inference-time RL in which an LLM improves by accumulating trajectory-reward pairs in its context, with the reward acting as a learning signal. This framing poses a central question: is ICRL genuine inference-time RL, or better understood as a form of in-context learning (ICL)? We investigate this in the context of direct ICRL, where the model directly uses trajectory-reward pairs. Through controlled experiments on three benchmarks across six models, we find that the reward is read, but its effect is small; randomizing or removing the reward leaves the improvement curve almost unchanged. Trajectories drive improvement, but not through their semantic content: shuffled or corrupted trajectories work as well as real ones. These patterns closely mirror those known in ICL, suggesting that direct ICRL is better understood as a special case of ICL than as inference-time RL. This reframing has implications for ICRL memory design: ICL factors such as input distribution and demonstrations may matter more than RL elements such as reward shaping and exploration.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a promising approach to enhance Instruction Following (IF) capabilities of large language models (LLMs). However, RLVR for instruction following remains sample-inefficient and prone to reward hacking, where LLMs exploit verification shortcuts rather than fulfilling the core intent. To address these challenges, we frame RLVR for instruction following as an integrated environment that unifies dynamic task generation and robust training. We introduce Instruction Following Decorator (IFDecorator}, a framework coupling difficulty adaptation, intent alignment, and hack diagnostics. It features: (1) a cooperative-adversarial data flywheel that yields challenging yet solvable tasks for sample-efficient training; (2) IntentCheck, a gating module to mitigate reward hacking; and (3) TripWires, a proactive diagnostic tool for eliciting hacking behaviors and quantifying their prevalence. Extensive experiments show our Qwen2.5-32B-Instruct trained with IFDecorator using only 3,625 training examples achieves 87.43% on IFEval (outperforming GPT-4o) and improves FollowBench by 4.2%, while preserving general capabilities. Diagnostics confirm that our method effectively reduces reward hacking. The approach generalizes across model architectures and scales. We will release code and data for future research.

Post-training has become essential for adapting large language models (LLMs) to complex downstream behaviors, including instruction following, preference alignment, and multi-step reasoning. Reinforcement learning with verifiable rewards (RLVR) has recently emerged as a particularly effective post-training paradigm for improving reasoning capabilities, with critic-free algorithms such as GRPO and GSPO enabling scalable optimization. However, RLVR post-training with full fine-tuning (FFT) requires substantial GPU memory and incurs high training costs. Although parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), effectively reduce computational costs, they often suffer from a noticeable performance gap compared to full fine-tuning in post-training for complex reasoning tasks. In this paper, we propose Hybrid-LoRA, an efficient hybrid post-training framework that selectively applies full fine-tuning to a small subset of modules less suited to low-rank adaptation, while adapting the remaining components with LoRA. We introduce a novel Hybrid-LoRA Score to rank candidate modules according to their sensitivity to low-rank adaptation under a fixed parameter budget. Experiments show that Hybrid-LoRA closely matches full fine-tuning performance under a 10\% full fine-tuning module budget, with the remaining candidate modules adapted by LoRA, consistently outperforming four state-of-the-art PEFT post-training baselines, achieving improvements of up to 5.65\% and on average 4.36\% over the best baseline.

While reinforcement learning (RL) successfully enhances reasoning in large language models, its role in fostering compositional generalization (the ability to synthesize novel skills from known components) is often conflated with mere length generalization. To this end, we study what RL post-training teaches models about skill composition and how the composition structure affects the skill trasnfer. We focus on the \countdown{} task (given $n$ numbers and a target, form an expression that evaluates to the target) and analyze model solutions using expression trees, where each subtree corresponds to a reusable subtask and thus can be viewed as a ``skill.'' Tracking tree shapes and their success rates over training, we find: (i) out-of-distribution (OOD) generalization to larger $n$ and to unseen tree shapes, indicating compositional reuse of subtasks; (ii) the order of skill acquisition depends upon the composition structure---models master shallow balanced trees (workload is balanced between subtasks) before deep unbalanced ones, with persistent fragility on right-heavy structures for a fixed composition depth. Our diagnostic reveals what is learned, in what order, and where generalization fails, clarifying how RL-only post-training induces OOD generalization beyond what standard metrics such as pass@k reveal. To study the implication of our findings for more real-world tasks, we present a preliminary study on formal theorem proving.

Reward modeling is not only a prediction problem: in KL-regularized policy optimization, the learned reward is exponentiated to define the deployed policy, so downstream value depends on errors in reward-tilted regions. We study this feedback in a Gaussian single-index model with $ r^* (x)=\sigma^* (\langle\theta^* ,x\rangle)$ and $x\sim\mathcal N(0,I_d)$. We analyze a two-stage neural reward model that first learns the hidden direction $\theta^*$ from reward-weighted samples and then fits the readout layer by weighted ridge regression. Exponential reward weighting changes the Hermite signal available to the first layer; for any feature-learning temperature $\beta_1$ above a dimension-free $O(1)$ threshold, a constant fraction of neurons recover the hidden direction, with weak-recovery complexity governed by the generative exponent. After feature recovery, we derive tilted-policy value-gap bounds for an idealized label-weighted fit with weights $e^{y/\beta_2}$ and a more practical surrogate-weighted fit with weights $e^{r_{a_0}(x)/\beta_2}$. Keeping the $\beta_2$-dependence explicit yields an admissible set of deployment temperatures, balancing the gain from lowering $\beta_2$ against the learning cost amplified by exponential weighting; in the surrogate-weighted case, proxy-dependent factors shrink this admissible set.

Learning accurate value functions plays a decisive role for reinforcement learning (RL) agents to solve long-horizon, complex tasks. Conventional temporal-difference (TD) learning objectives suffer from value-estimation bias that accumulates over the horizon, while extended-horizon modeling methods, such as $n$-step TD backups and Q-chunking, adopt a rigid, fixed-horizon value-modeling recipe that is often not flexible enough to capture complex value structures in long-horizon, multi-stage tasks. In this paper, we show that enabling value updates with dynamic horizon composition can yield a strong offline policy learning scheme. Our method, _Horizon Adaptive Offline Policy Learning via **VA**lue **ST**itching_ (**_VAST_**), replaces fixed-horizon backups with recursive, horizon-adaptive value composition. Its key ingredient is to couple value optimization with a future state- and horizon-length-conditioned **_auxiliary value function_** that is learned through direct data supervision, and a **_stitching policy_** that optimally selects the reward-maximizing horizon length and future sub-goal to achieve horizon-adaptive value stitching. This design enables direct estimation and compositional “stitching” of variable-length returns grounded in actionable sub-goal states, providing an accurate and greedily exploitable value-supervision signal for offline policy optimization. Across 50 tasks on OGBench, _VAST_ outperforms fixed-step, extended-horizon methods, and generative-value offline RL baselines, achieving strong performance particularly in high-complexity, long-horizon decision-making tasks.

Vision-language models (VLMs) show strong multimodal capabilities but still struggle with fine-grained vision-language reasoning. We find that long chain-of-thought (CoT) reasoning exposes diverse failure modes, including perception, reasoning, knowledge, and hallucination errors, which can compound across intermediate steps. However, most existing data for reinforcement learning with verifiable rewards (RLVR) does not involve complex reasoning chains grounded in visual evidence throughout, leaving these failure modes understressed during training. We therefore propose HopChain, a scalable framework that constructs an out-of-distribution (OOD) proxy task: OOD only in query format, while strengthening foundational abilities shared with downstream benchmarks to support generalizable gains. Concretely, HopChain synthesizes multi-hop vision-language reasoning chains, each a logically dependent chain of instance-grounded hops where earlier hops establish the instances, sets, or conditions needed for later hops, ending in an unambiguous numerical answer for verifiable rewards. In experiments, we train Qwen3.5-35B-A3B and Qwen3.5-397B-A17B under two RLVR settings (with and without HopChain's multi-hop data) and compare them across 24 benchmarks spanning STEM and Puzzle, General VQA, Text Recognition and Document Understanding, and Video Understanding. Although the multi-hop data is not designed for any specific benchmark, it improves 20 of 24 benchmarks on both models, indicating broad and generalizable gains. Consistently, replacing full chained queries with half-multi-hop or single-hop variants reduces the average score across five representative benchmarks from 70.4 to 66.7 and 64.3, respectively. Improvements are particularly substantial in long-CoT and ultra-long-CoT regimes, peaking at more than 50 accuracy points in the ultra-long-CoT regime. These results show that OOD proxy tasks for long-chain visual reasoning are a scalable RLVR supervision source for generalizable VLM reasoning.

Prior Laplacian-based option-discovery methods build options from 0-form Laplacians defined on nodes, where the resulting eigenvectors induce routes toward isolated boundaries or extrema, thereby skewing visitation toward specific states. In contrast, kernel eigenvectors of the 1-form Hodge Laplacian represent interpretable, topology-aware edge flows and naturally capture cyclic behaviour known as harmonic flows. These kernel eigenvectors are particularly attractive because they define zero-divergence edge flows, offering a natural solution for non-isolated visitation to sources and sinks in regions in which they are active. Despite this appeal, harmonic flows have not previously been explored as a basis for option discovery in reinforcement learning. This is what we do here. We first introduce harmonic-flow skills and show that, on state graphs with tree-like “dead-end” appendages, purely harmonic flows can become nearly inactive in those regions, which can limit coverage. To address this limitation, we propose quasi-harmonic flow options: skill-inducing edge flows that preserve harmonic cycles where harmonic activity is strong, while injecting drift into appendages using the lowest-magnitude non-harmonic eigenvector of the Hodge Laplacian, a low divergence solution, where harmonic activity is weak. Our approach is eigenvector-only, relying solely on spectral components of the Hodge Laplacian. We compare quasi-harmonic flow options against a harmonic-flow options baseline and state-of-the-art skill-discovery methods from the literature, demonstrating empirical gains and analytical improvements on hitting-time-based metrics.

We present a new operator-theoretic representation learning framework for offline reinforcement learning that recovers the directed temporal geometry of a controlled Markov process from hitting time observations. While prior art often produces symmetric distances or fails to satisfy the triangle inequality, our framework learns a Hilbert-space displacement geometry where expected hitting times are realized as linear functionals of latent displacements. We prove that this representation exists under latent linear closure and is uniquely identifiable up to a bounded linear isomorphism. For finite-dimensional implementations, we show that global hitting-time error is bounded by one-step transition error amplified by the environment's transient spectral radius. Furthermore, we provide finite-sample guarantees accounting for approximation, statistical complexity, and trajectory-label mismatch. Derived from this theory, we curate Isomorphic Embedding Learning (IEL) as a new goal-agnostic foundation policy learning algorithm that anchors a HILP-style consistency objective with explicit hitting-time regression to ensure that the learned geometry reflects actual decision-time progress. This asymmetric and compositional structure enables robust graph-based multi-stage planning for long-horizon navigation. Our experiments demonstrate that IEL improves the state of the art of learning foundation policy policies from offline maze locomotion data.

Modern data-driven decision-making methods, such as imitation learning (IL) and reinforcement learning (RL), have achieved great success in solving many complex tasks. However, these methods often suffer from serious control instability and robustness issues when applied in real-world applications such as robotics and autonomous driving, posing notable challenges for their practical deployment. We argue that this instability issue stems largely from their limitations on solely supervising and optimizing zeroth-order actions (i.e., the action labels), failing to account for higher-order action dynamics and temporal consistency. In this paper, we show that simultaneously supervising both zeroth- and first-order actions can dramatically enhance policies' performance and control robustness. To achieve this, we introduce a novel and elegant loss scheme supported by formal theoretical guarantees that can equip any off-the-shelf policy model (e.g., deterministic, stochastic, or flow policies) with the capability for higher-order action supervision, without requiring any structural modifications. Moreover, our proposed method can serve as a lightweight plug-and-play module that seamlessly integrates with a broad spectrum of existing offline RL frameworks. Extensive evaluations on OGBench and D4RL demonstrate that our approach yields substantial performance and robustness improvements across a wide range of continuous control environments. Notably, our method can also enhance policies' out-of-distribution (OOD) generalization capability in the challenging low-data regime, making it an ideal tool in tackling many real-world control problems.

High-dimensional robotic reinforcement learning (RL) is often bottlenecked by inefficient exploration in large, redundant actuator spaces. We introduce Devyn (Developing synergies), a lightweight framework that learns a structured low-dimensional action parameterization directly during policy optimization. Unlike methods that require precomputed synergies or structured exploration in the full action space, Devyn starts from a randomly initialized latent-to-action map and shapes it online through task interaction. The policy acts strictly in a reduced latent space, while a learnable synergy matrix decodes latent actions into actuator commands. To make this evolving decoder stable and useful for policy learning, Devyn combines slow decoder evolution with structural regularization, reducing latent-action semantic drift while encouraging compact coordination patterns. Across diverse continuous-control benchmarks, including high-dimensional humanoid and overactuated musculoskeletal systems, Devyn improves upon standard SAC and achieves competitive or better performance than recent methods designed for high-dimensional robotic control. Ablations show that the gains do not come from a generic latent-action bottleneck alone, but rather from the slow, structured evolution of the synergy matrix. Our results suggest that learning the action parameterization itself can be a simple and effective route to scaling RL for high-dimensional robots.

We propose a distributional theory of how hypernymy---the ``is-a'' relation between general and specific concepts---is encoded geometrically in language representations. Starting from the empirically verified assumption that words closer on the WordNet hypernym graph co-occur more often, we characterize theoretically the spectrum of the resulting embedding Gram matrix of word2vec embeddings. Under mild positivity and decay conditions on the co-occurrence kernel, we prove that the leading eigenvectors first separate broad taxonomic branches and then progressively finer sub-branches, producing a \emph{hierarchical splitting geometry} with a coarse-to-fine spectral organization that mirrors the tree. We confirm these predictions in word2vec embeddings across many sampled WordNet subtrees, and show that the same signature extends strikingly well to Gemma 2B unembeddings. Our results indicate that hierarchical concept geometry in LLMs need not reflect a hierarchy-specific functional mechanism, but emerges from the spectral structure of pairwise word statistics.

Long-horizon visual search requires a vision-language policy to execute multi-step spatial actions over high-resolution images, where performance depends on the entire trajectory rather than any single decision. Existing supervision breaks at this horizon: outcome-only RLVR collapses many geometric decisions into a single delayed signal and discards the per-step geometry the environment already computes, while LLM-judged step rewards recover density at the cost of verifiability, scale linearly with trajectory length, and conflate distinct failure modes. We propose HSSP, a self-play framework in which a Hider adaptively weights challenging skeletons from a fixed trajectory pool and a Seeker learns to solve them through grounded spatial actions. Instead of relying on external judges, HSSP supervises every step with a geometry-verifiable process reward built from target-directed progress and exploration coverage, with trajectory length and spatial redundancy enforced as separate Constrained Markov Decision Process (CMDP) constraints. We pair HSSP with V-Trace, a trajectory-level companion dataset that augments skeleton seeds from established benchmarks with target-region annotations and explicit search-regime labels, ready for any trajectory-level training framework. The reward design admits closed-form guarantees on backtracking, find-versus-explore separation, and policy invariance, providing theoretical backing for the observed gains. Extensive empirical results on eleven challenging multimodal benchmarks, supported by controlled ablations and qualitative case studies, show that HSSP consistently outperforms advanced baselines and transfers zero-shot to OOD tasks.

Reinforcement learning with verifiable rewards (RLVR) has become a powerful paradigm for improving language models on reasoning-intensive tasks, but its effectiveness is often limited by exploration. For example, models often fail on hard problems, leaving little useful reward signal. External expert traces offer a natural source of guidance, yet they may also expose reward-relevant content along the critical path to the verifier target, such as final answers, intermediate values, executable implementations, or answer-related entities. This content can create an unintended reward hacking channel, allowing the policy to obtain reward by copying the trace rather than learning the underlying reasoning or agentic behavior. Existing guided-RL methods reduce this risk by using partial trajectories, but they mainly control how much expert information is shown heuristically rather than which parts should be hidden. To this end, we propose Semantic Masked Expert Policy Optimization (SMEPO), a fine-grained semantic masking strategy for expert-guided RLVR. Instead of truncating traces coarsely or revealing them unchanged, SMEPO masks reward-relevant semantic spans along the critical path while preserving the expert’s decomposition, plan, and procedural structure. This turns hard problems from reasoning from scratch into a fill-in-the-blank process: the policy can follow the expert’s problem-solving route, but must still reconstruct the missing values, code, or entities by itself. SMEPO is simple to apply and requires no changes to the reward function or RL objective. Across diverse domains, including math, code, and agentic search, SMEPO improves accuracy by up to 3.2 points over GRPO and reduces training time by up to 4.2$\times$.

Video generation models have achieved remarkable visual fidelity and have strong potential to become general-purpose world simulators. Despite this progress, they still fail to generate videos which adhere to laws of physics. The problem becomes even more apparent in realistic settings where multiple physical principles must work together within the same video; for example, "a balloon floating upward while steam rises from a pot" requires buoyancy and fluid dynamics to unfold coherently and simultaneously. Yet existing methods largely ignore multi-principle interactions, focusing on a single principle per video. We propose HiPhy (Hierarchical Physical Alignment), a reinforcement learning framework that grounds video generation in physical laws through a dual-level objective: locally enforcing the temporal dynamics of individual physical principles, and globally ensuring the physical and semantic coherence of the entire scene. To support multi-principle generation, we construct a 50K-prompt dataset and introduce a prompt benchmark MultiPhyBench, spanning a diverse range of co-occurring physical events. Our experiments show that HiPhy significantly outperforms prior methods and baselines, improving physical commonsense and semantic alignment significantly across various benchmarks, with the largest gains on scenes involving multiple concurrent physical principles where competing methods degrade most sharply.