NeurIPS 2026 论文列表

Big goals are hard to achieve all at once; breaking them into small steps is wiser. We present Trust Region Policy Distillation (TOP-D), which transforms the notoriously unstable, high-variance On-Policy Distillation (OPD) into a stable training paradigm by dynamically constructing a proximal teacher. Theoretically, we establish a rigorous framework demonstrating that TOP-D inherently controls gradient variance. By providing a formal global convergence analysis alongside a monotonic improvement bound, we mathematically formalize the reliability and stability of the overall training dynamics. Empirically, TOP-D dramatically enhances training stability, sample efficiency, and final performance on mathematical reasoning tasks. More importantly, TOP-D introduces zero additional computational overhead, positioning itself as a promising alternative to the well-established OPD paradigm.

Understanding the benefits, risks, and impacts of general-purpose AI assistants requires looking at real conversations---not curated benchmarks or surveys. Yet AI use narratives rely on limited proprietary reports, or on few data sources, fragmented across platforms, models, and time. We introduce the AI Observatory, a public measurement platform that aggregates the most comprehensive set of real AI conversation sources, and develop a common taxonomy of 145 features, spanning function, topic, sensitive use, interaction style, multi-turn dynamics, and conversation structure. Across 23,158 conversations and 85,633 turns, we find that real AI use is highly heterogeneous: sources differ substantially in tasks, structures, and sensitive-use distributions, with no single source that generalizes. We further show that occupational summaries such as Anthropic's Clio capture an important but incomplete slice of use: 48\% of the conversations we analyze would be filtered out for being non-occupational, omitting substantial personal, social, cultural, and safety-relevant interaction. Real AI use is also not static: we find tokens and turns per conversation all rise sharply between 2023 and 2025, and we expose how model variants within the same developer support distinct usage regimes. The Observatory provides a new taxonomy, reusable annotation tools, and a public platform for reproducible measurement of how AI assistants are really used and evolving in the wild: \url{https://project-ai-observatory.vercel.app/}.

Reinforcement learning from verifiable rewards improves reasoning by reinforcing sampled solutions that receive positive outcomes, but group-relative methods such as GRPO become silent on hard prompts where every sampled rollout is wrong. These zero-reward prompts are often the most informative failures: each comes with a verified reference solution, yet uniformly imitating the full trace mixes core reasoning decisions with routine algebra, formatting, and surface wording. We propose **Selective Off-Policy Reference Tuning with Plan Guidance (SORT)**, an auxiliary repair update that leaves GRPO rollout generation unchanged. For each failed prompt, SORT extracts a reference-derived reasoning plan and scores every reference token twice under the same model, once with only the problem and once with the plan added to the context. Tokens whose probabilities rise under plan conditioning are treated as structurally informative and receive larger Dynamic Fine-Tuning updates, while tokens outside the model's support remain controlled by the base probability. We formalize this mechanism with a ground-truth plan model and prove that the SORT weight approaches an oracle structural-token weight as the extracted plan better approximates the true plan-conditioned distribution. Across three instruction-tuned backbones and eight in-distribution and out-of-distribution reasoning benchmarks, SORT consistently improves over GRPO and guidance-based baselines, with the largest gains on the weakest model where zero-reward failures are most frequent.

High-quality teleoperation datasets are costly to collect, particularly for hard tasks. We observe that many tasks exhibit directional asymmetry: completing the forward hard task is difficult, whereas reversing it by relaxing or disrupting the environment is comparatively easy. This suggests that reversed easy-task trajectories can serve as a scalable supervision signal for the hard task, reducing the cost of manual demonstration collection. However, reversed data can be noisy, and directly training on it may yield suboptimal policies. To enable largely automated acquisition and effective use of reversed data, we propose a teleoperation-cost effective framework for hard policy learning via temporal reversal of easy tasks, consisting of three key components: a closed-loop data collection pipeline that alternates between hard-task and easy-task policies to autonomously reset the environment and generate diverse trajectories; a hierarchical data refinement pipeline that temporally inverts easy-task rollouts and filters low-quality motion using kinematic priors and a critic-guided advantage filter; and an iterative policy learning method that trains the hard-task policy using both initial reversed easy-task demonstrations and the filtered reversed data in a continuous online learning loop. By combining automated collection, hierarchical refinement, and iterative learning, our method enables scalable, reliable training of complex, high-precision manipulation tasks. Across two simulated benchmarks and real-robot experiments, we demonstrate that our method improves hard-task success rates with higher data efficiency and more stable training compared to reversal-based and reinforcement-learning baselines, without requiring extensive hard-task teleoperation.

On-policy distillation (OPD) improves language-model post-training by letting the student learn from its own sampled trajectories under teacher supervision, aligning the training process with inference-time behavior. However, the rollout policy in OPD is usually treated as fixed: each round samples directly from the current student. We show that this choice can create a compounding teacher-student mismatch, because it requires the teacher to provide effective supervision on whichever trajectories the student samples, including ones less reachable by the teacher itself. On such trajectories, teacher supervision can become noisy or even misleading, leading to undesirable student patterns such as repetition. Moreover, since the updated student in turn defines the rollout policy of subsequent rounds, these patterns persist and potentially amplify over training. This failure mode suggests that improving OPD requires treating the rollout policy itself as an optimization target, rather than only modifying the distillation objective. We propose *Mixture Proposal On-Policy Distillation* (MPOPD), which derives a student-anchored, teacher-guided rollout distribution from a local trust-region objective based on relative teacher support. The resulting target is a geometric mixture of the student and teacher policies, and we introduce a proposal policy that approximates this mixture, enabling rollout generation in a single forward pass. Experiments on 16 math, knowledge, and medical reasoning benchmarks show consistent gains over OPD and recent variants. On math reasoning, our method improves average accuracy by 6.91 over OPD on Qwen3-4B-Base while reducing average generation length by 41\%, suggesting that rollout policy design provides an effective way to improve OPD.

We study policy regret minimization in partially observable Markov games (POMGs) between a learner and a strategic adaptive adversary who adapts to the learner's past strategies. We develop a model-based optimistic framework that operates on the learner-observable process using *joint* MLE confidence set and introduce an Observable Operator Model-based causal decomposition that disentangles the coupling between the world and the adversary model. Under multi-step weakly revealing observations and a bounded-memory, stationary and posterior-lipschitz adversary and planner stability, we prove an $\mathcal{O}(\sqrt{T})$ policy regret bound. This work advances regret analysis from Markov games to POMGs and provides the first policy regret guarantee under imperfect information against an adaptive opponent.

While there is an extensive body of research analyzing policy gradient methods for discounted cumulative-reward MDPs, prior work on policy gradient methods for average-reward MDPs has been limited, with most existing results restricted to ergodic or unichain settings. In this work, we first establish a policy gradient theorem for average-reward multichain MDPs based on the invariance of the classification of recurrent and transient states. Building on this foundation, we develop refined analyses and obtain a collection of convergence and sample-complexity results that advance the understanding of this setting. In particular, we show that the proposed $\alpha$-clipped policy mirror ascent algorithm attains an $\epsilon$-optimal policy with respect to positive policies.

Off-policy reinforcement learning (RL) tries to learn the value of a target policy while data comes from a different behavior policy. This often requires trading off bias for variance or stability. One way to trade these off is to linearly interpolate between the behavior and target policies, but this represents only one point in a broader space of interpolation mechanisms. We introduce \emph{excursion policies}--a structured alternative where the agent follows the target policy for a multi-step ``excursion'' before reverting to the behavior policy. We derive dynamic programming results for these policies, establishing the theoretical basis for their use in RL. We further develop provably sound off-policy temporal-difference (TD) algorithms for excursion value estimation. We find that excursion policies provide an efficiency-stability trade-off comparable to mixture policies in continuous control benchmarks, and our analysis illustrates conceptual differences in their behavior.

Deployed language models are increasingly expected to remove private, copyrighted, or hazardous content without full retraining. Existing unlearning methods remain brittle because they usually optimize a fixed forgetting objective with a global update schedule: weak updates leave paraphrased knowledge intact, while stronger updates can damage retained capabilities or nearby facts. We identify this failure mode as update homogeneity. To address it, we introduce METAFORGET, a meta-optimization framework that treats unlearning as learning an audit-driven update policy rather than designing another standalone forgetting loss. A lightweight controller maps memorization, paraphrase-sensitivity, and signed forget-retain conflict signals to span weights, layer gates, step sizes, and retain-gradient projection strengths. The controller is trained with a bilevel objective that differentiates through short LoRA unlearning trajectories and evaluates the resulting model on held-out forget, retain, locality, and leakage audits. This policy can instantiate NPO- or RMU-style losses while changing where and how strongly they act. Across TOFU, WMDP, and MUSE-style sequential deletion settings, METAFORGET improves the forget-retain-robustness frontier over strong baselines, moves loss-based membership-inference diagnostics closer to chance, and better preserves utility under repeated requests. The results support a view of practical LLM unlearning as audit-driven policy learning: the central object is not only the forget loss, but the localized update that must pass the deletion audit.

Offline reinforcement learning aims to learn a policy solely from fixed datasets, which often contain multimodal action distributions. Flow policies can naturally represent such multimodal behaviors, but learning an efficient one-step flow policy remains challenging: standard value guidance often leads to mode collapse or exploits overestimation bias in out-of-distribution regions. To address this, we introduce One-step Flow policy via Optimal Transport (OptiFlow), a framework for one-step flow policy learning as a structured sample-allocation problem. OptiFlow jointly trains a value-aware reference flow policy and an efficient one-step policy, coupling their action samples through state-wise entropic optimal transport. For each state, critic-estimated values define the priority of distillation target actions, while the action-distance cost ensures geometrically compatible pairings. By avoiding direct critic maximization, our transport-guided approach enables in-distribution exploitation by anchoring the one-step policy to high-value, dataset-supported modes without the risk of out-of-distribution divergence. Experimental results demonstrate that OptiFlow effectively captures optimal multimodal behaviors and achieves strong performance across diverse offline RL benchmarks.

Reliable progress in offline policy learning depends on a small set of methodological practices, including careful reporting, well-tuned baselines, and evaluation across diverse conditions. Prior work has noted that results can be sensitive to reporting choices, hyperparameter tuning, and dataset properties independently, but these sources of variability have not been systematically investigated at the scale needed to understand how they shape conclusions. To address this gap, we present a large-scale empirical study of offline reinforcement and imitation learning, training over 145,000 policies across 114 datasets. At this scale, no algorithm dominates: aggregate performance across top methods is often close, but the leaders differ substantially across environments. We find that proper hyperparameter tuning frequently reshuffles perceived algorithm rankings; and simple baselines, including behavior cloning, are often stronger than is commonly assumed after extensive tuning. We also study hyperparameter transfer and sensitivity across environments, identifying a simple strategy that generalizes well. From these analyses we distill practical recommendations, and release JumpStart: a resource suite of trained models, per-model hyperparameter and reward data, strong baselines across all environments, and a website to make retrieval and analysis trivial. Together, these resources aim to make offline policy learning research more reliable and to open new directions for work beyond the scope of this study.

We study the global convergence of policy gradient for infinite-horizon entropy-regularized Markov decision processes (MDPs) with continuous state and action spaces. We consider log-linear softmax policies with linear function approximation, which extend the tabular softmax parameterization while retaining a tractable policy class. Under $Q^\pi_\tau$-realizability for the regularized state-action value function, we first establish a non-uniform Polyak--Łojasiewicz (PŁ) inequality. The non-uniformity arises through degeneracy of constants associated with the policy geometry, namely the Fisher information matrix or an uncentered feature covariance matrix. We then identify two feature regimes under which this non-uniform constant can be bounded along the gradient flow. For full-affine-span features, we prove radial unboundedness of the KL regularizer and show that the smallest eigenvalue of the Fisher information matrix remains bounded below by an initialization-dependent positive constant. For simplex-valued features, we prove an analogous radial unboundedness result in the subspace orthogonal to the all-ones vector and obtain a uniform lower bound for the smallest eigenvalue of the uncentered covariance matrix. These results imply global linear convergence of the regularized objective along the gradient flow, i.e. suboptimality decaying as $\mathcal{O}(e^{-Ct})$ for some $C>0$. Our analysis extends the global convergence theory of entropy-regularized softmax policy gradient beyond the tabular setting of Agarwal et al. [2020], Bhandari and Russo [2024], Mei et al. [2020].

World models aim to improve robotic decision making by predicting the consequences of actions. However, in practice, their predictions often become unreliable once the robot encounters states outside the training distribution, limiting their effectiveness at deployment. We observe that execution itself provides a natural but underutilized signal: after each action, the robot directly observes the true next state, revealing the mismatch between predicted and actual outcomes. Building on this insight, we propose $\textbf{feedback world model}$, a new paradigm that closes the loop between prediction and observation at inference time. Instead of treating the world model as a static open-loop predictor, our method maintains a lightweight feedback state that is updated online to iteratively correct future predictions, compensating for model errors using real-time observations without additional training data or parameter updates. We show that this process can be interpreted as a latent-space observer and admits convergence guarantees under mild conditions. We further introduce action-aware guidance to better translate corrected predictions into control by emphasizing action-controllable components while suppressing irrelevant variations. Experiments on LIBERO-Plus, Robomimic, and real-world manipulation tasks demonstrate that our method substantially improves both prediction accuracy and policy performance under distribution shift. In particular, it reduces world model prediction error by up to 76.4\% and improves out-of-distribution (OOD) success rate by 30\%. These results show that incorporating real-time feedback at inference time provides a simple yet powerful alternative to static world modeling.

Maximum entropy reinforcement learning (MaxEnt-RL) enables robust exploration, yet practical implementations often restrict policies to simple Gaussians. While recent approaches incorporate expressive generative policies via importance-weighted supervised learning, they are prone to importance weight collapse, which limits their scalability in high-dimensional action spaces. Our key insight is to mitigate this limitation by localizing the sampling region, avoiding the weight degeneracy induced by importance sampling over the entire action space. To instantiate this insight, we introduce **FLAG** (**F**low policy with **L**atent-**A**ugmented **G**uidance). FLAG augments the state space with a flow latent variable and optimizes a provably consistent proxy MaxEnt-RL objective. We empirically demonstrate that FLAG enables expressive policy optimization with limited importance samples and scales to high-dimensional control tasks. Furthermore, FLAG achieves state-of-the-art performance across challenging benchmarks.

Reinforcement-learning-based post-training for large language models vastly improves capabilities on verifiable tasks however it suffers from training instability and diversity collapse. As a solution, many new advantage functions have emerged. Each method simultaneously changes which problems receive gradient, the balance between positive and negative updates, and the overall gradient scale making their comparison difficult. We propose a unifying framework that decomposes each function by their induced positive and negative gradient mass. Analyzing policy performance and the geometry of model weights reveals: symmetric methods are better but slow and asymmetric methods are fast. Combining the advantages of both, we build the FADE: Focal Adaptive Dynamic Entropy advantage which is fast, diverse and accurate at various model scales (7B, 32B).

Policy learning from observational data is fundamental to applications ranging from personalized medicine to welfare-program targeting, but standard approaches optimize empirical performance and may fail when the deployment distribution differs from the training distribution. We study \emph{distributionally-robust policy learning}: given observational data $\\{(X_i,W_i,Y_i)\\}_{i=1}^n$ with binary treatment $W$, we seek a policy $\pi \in \Pi$ that maximizes the worst-case policy value over an $f$-divergence ball of radius $\rho$ around the data-generating distribution. We propose an estimator built on cross-fitted doubly-robust scores for the conditional treatment effect, paired with a DRO solver that handles \texttt{KL}, $\chi^2$, and \texttt{CVaR} ambiguity sets. We establish two convergence guarantees. First, the estimated policy attains the standard $\widetilde{O}(n^{-1/2})$ rate on its DRO regret, and inherits the doubly-robust property from its non-DRO counterpart -- the rate continues to hold whenever either the outcome model or the propensity model is consistently estimated, even at slow nonparametric rates. We then propose a sample-splitting bias-correction algorithm derived from the Lagrangian dual, and show that under a Tsybakov-style margin condition on the worst-case treatment-effect contrast, the estimator attains a fast rate of $\widetilde{O}(n^{-(1+\alpha)/(2+\alpha)})$ , interpolating between $\widetilde{O}(1/\sqrt{n})$ and $\widetilde{O}(1/n)$ as the margin parameter $\alpha$ ranges over $[0,\infty)$. To our knowledge, this is the first work to incorporate a Tsybakov margin condition into the analysis of distributionally-robust policy learning, and the first to obtain fast rates in this setting. Empirically, on semi-synthetic experiments calibrated to a welfare-to-work program, distributional robustness yields meaningful improvements over empirical welfare maximization when training and test distributions differ structurally, while remaining competitive when no shift is present.

While recent work in Reinforcement Learning with Verifiable Rewards (RLVR) has shown that a small subset of critical tokens disproportionately drives reasoning gains, an analogous token-level understanding of On-Policy Distillation (OPD) remains largely unexplored. In this work, we investigate high-loss tokens, a token type that—as the most direct signal of student-teacher mismatch under OPD's per-token KL objective—should progressively diminish as training converges according to existing studies; however, our empirical analysis shows otherwise. Even after OPD training reaches apparent saturation, a substantial subset of tokens continues to exhibit persistently high loss; these tokens, which we term Rock Tokens, can account for up to 50\% of the tokens in generated outputs. Our investigation reveals two startling paradoxes. First, despite their high occurrence frequency providing a disproportionately large share of total gradient norms, Rock Tokens themselves remain stagnant throughout training, resisting teacher-driven corrections. Second, through causal intervention, we find that these tokens provide negligible functional contribution to the model's actual reasoning performance. These findings suggest that a vast amount of optimization bandwidth is spent on structural and discourse residuals that the student model cannot or need not internalize. By deconstructing these dynamics, we demonstrate that strategically bypassing these "stumbling blocks" can significantly streamline the alignment process, challenging the necessity of uniform token weighting and offering a more efficient paradigm for large-scale model distillation.

Learning from multiple imperfect oracles is a practical way to improve reinforcement learning (RL) under limited interaction. However, most contemporary robust multi-oracle methods are on-policy and equire frequent fresh rollouts, which limits sample efficiency in domains where data collection is expensive. We propose Oracle-Aggregated Policy Improvement (OPI), an off-policy framework for policy improvement with multiple suboptimal oracles. OPI augments the oracle set with the learner, constructs an oracle-guided behavior policy through candidate-action pooling and source-aware critic scoring, and trains the learner using a generalized advantage-weighted behavior cloning objective combined with deterministic policy gradients. The method is motivated by a tractable proxy for the max-advantage objective and a policy-improvement relative to a strong oracle baseline. We evaluate OPI on diverse tasks across MetaWorld, the DeepMind Control Suite, and customized Box2D environments with heterogeneous oracles. Across sparse-reward manipulation, dense-reward locomotion, and customized control settings, OPI is strongest in sparse-reward and complementary-oracle regimes and remains competitive with representative baselines in dense-reward environments. Additional analyses show that oracle composition is most beneficial when oracle skills are complementary and that estimating oracle-specific action values is particularly helpful in sparse-reward regimes

RLVR and OPD have become standard paradigms for post-training. We provide a unified analysis of these two paradigms in consolidating multiple expert capabilities into a single model, identifying capability loss in different ways: mixed RLVR suffers from inter-capability divergence cost, while the pipeline of first training experts and then performing OPD, though avoiding divergence, fails to fully absorb teacher capabilities due to large behavioral pattern gaps between teacher and student. We propose **Co-Evolving Policy Distillation (CoPD)**, which encourages parallel training of experts and introduces OPD during each expert's ongoing RLVR training rather than after complete expert training, with experts serving as mutual teachers (making OPD bidirectional) to co-evolve. This enables more consistent behavioral patterns among experts while maintaining sufficient complementary knowledge throughout. Experiments validate that CoPD achieves all-in-one integration of text, image, and video reasoning capabilities, outperforming strong baselines such as mixed RLVR and MOPD, and even surpassing domain-specific experts.