NeurIPS 2026 Papers

Context-based offline meta-reinforcement learning (COMRL) aims to learn adaptable policies entirely from static datasets by inferring latent task representations from small contexts. Most COMRL methods rely on immediate, per-step reward signals. This can be generalized by aggregating rewards into ``bags'' and being delivered after a sequence of actions (where bag size $n=1$ is the per-step case). We formally prove that under Bagged Rewards, existing per-transition COMRL methods suffer an information-theoretic collapse ($\mathcal{O}(1/n)$), causing their task inference to degrade to behavior cloning. To solve this, we introduce SpectralMeta, an algorithm that fully decouples dynamics representation from task inference. By pre-training an energy-based spectral representation on reward-free transitions, we recast task identification as a closed-form Bayesian linear regression over bagged rewards. SpectralMeta comes with finite-sample bounds on reward recovery and policy sub-optimality. In continuous-control meta-RL benchmarks, SpectralMeta maintains robust task inference and competitive returns even when rewards are reduced to a single episode-level scalar, a regime where existing baselines fail to recover task-relevant structure.

In this work, we study transformer attention through the lens of spectral geometry and operator theory. We view each attention head as a functional map between Hilbert spaces of functions on the token sequence and derive a \emph{Token Difference Operator}, whose spectral structure controls how token-space information is routed to the output. We show that standard Euclidean spectra are structurally biased by sinks, conflating mass concentration with genuine routing capacity. By recasting token space in the intrinsic probability geometry induced by attention, the token difference spectrum disentangles sink effects from routing capacity and provides a spectral description of the dimensionality of the head output. This yields a unified framework for analyzing attention maps, explaining sinks, routing collapse, and output dimensionality within a single operator-theoretic framework. In practice, by grounding attention heuristics in spectral geometry, we develop a novel uncertainty quantification method for LLMs that provides a theoretically sound alternative to traditional, probability-based estimators, with the strongest gains on long-form summarizations

Recent online reinforcement learning has substantially improved image editing quality. However, existing Flow-GRPO-style methods usually rely on a single whole-image reward, which makes fine-grained editing optimization difficult. We observe that a key obstacle in image editing is this spatial uniformity assumption: a whole-image reward cannot distinguish how different spatial regions contribute to image quality. To address this issue, we propose **SpatialFlow-GRPO**, a training framework that makes spatial reward feedback more fine-grained. The framework converts region-aware rewards into semantic-region-level optimization signals and aligns region advantages with the corresponding latent positions during policy updates. We also train a region-aware reward model, **SFReward**, construct **SFReward-14K** with region-annotated editing samples, and introduce **MultiEditBench** to evaluate multi-region editing ability. On OmniGen2 and FLUX.2-klein-4B, SpatialFlow-GRPO outperforms Flow-GRPO on GEdit-Bench, ImgEdit-Bench, and MultiEditBench. The results show that SpatialFlow-GRPO can turn local feedback into more accurate update signals and improve editing quality.

Multimodal large language models (MLLMs) excel at visual interpretation but fail on spatial reasoning tasks that humans solve reliably. Existing benchmarks evaluate these models as black boxes, limiting their ability to identify the underlying causes of lower performance: when a model fails a spatial reasoning task, it remains difficult to ascertain whether the hurdle is perceptual, such as recognizing object boundaries, or cognitive, such as reasoning about occlusion to infer hidden geometry. We introduce Spatial-IQ, a hierarchical diagnostic framework that decomposes object counting in stacked 3D structures into 9 perceptual and cognitive sub-tasks organized by the developmental stages of human spatial cognition, with mental rotation as an additional target probe. Using NVIDIA Isaac Sim, we procedurally generated a diverse dataset of roughly 80,000 stacked 3D structures with per-task ground truth. We evaluate models across three output formats (free-response text, multiple-choice images, and image editing) alongside a human baseline. The Spatial-IQ framework shows that top-performing models often succeed at the target task (object counting) without succeeding on the lower-level sub-tasks intended to support it, and that models differ in how much of these hierarchical chains they preserve, often revealing shortcut behavior that raw target-task accuracy alone would obscure. Finally, we demonstrate that training models with chain-of-thought (CoT) supervision over our hierarchical sub-tasks, combined with reinforcement learning with verifiable rewards, significantly improves both spatial consistency across sub-tasks and target-task accuracy, supporting the value of the proposed decomposition as both a diagnostic tool and a training signal.

Recent advancements in zero-shot reinforcement learning (RL) have facilitated the extraction of diverse behaviors from unlabeled, offline data sources. In particular, forward-backward algorithms (FB) can retrieve a family of policies that approximately solves any standard RL problem (with additive rewards, linear in the occupancy measure), given sufficient capacity. While retaining zero-shot properties, we tackle the greater problem class of RL with general utilities, in which the objective is an arbitrary differentiable function of the occupancy measure. This setting is strictly more expressive, capturing tasks such as distribution matching or pure exploration, which may not be reduced to additive rewards. We show that this additional complexity can be captured by a novel, maximum entropy (soft) variant of the forward-backward algorithm, which recovers a family of stochastic policies from offline data. When coupled with zero-order search over compact policy embeddings, this algorithm can sidestep iterative optimization schemes, and optimizes general utilities directly at test-time. Across both didactic and high-dimensional experiments, we demonstrate that our method retains favorable properties of FB algorithms, while also extending their range to more general RL problems.

Reinforcement learning with verifiable rewards (RLVR) has greatly advanced large reasoning models (LRMs), but it requires timely training on a huge fully-annotated dataset. To this end, data-efficient RLVR methods have been widely studied from two perspectives: (i) data selection methods identify a small subset of "golden" samples that yield near-full-data performance, but they rely on a pre-existing pool of labeled data. (ii) unsupervised RLVR methods train the model using its own internal supervision signals on large-scale unlabeled data, yet they exhibit suboptimal performance. Accordingly, we investigate the "*pick in the dark*" setup for RLVR, which aims to select, without prior supervision, unlabeled samples that are most beneficial for training and worthy of annotation. Through systematic analysis, we demonstrate that smart picks hinge on a well-calibrated uncertainty estimator to enable strategic partitioning of data for adaptive training regimes. Building on this insight, we propose **PivotTrace**, a three-way data triage framework that leverages attention dynamics to trace *metacognitive pivots* during reasoning. By precisely quantifying uncertainty through pivot density, PivotTrace achieves automated data routing to synergistically maximize both annotation and training efficiency. Empirically, PivotTrace surpasses the fully supervised LRM with only **29.3%** annotated samples and **2.75**$\times$ faster convergence.

Memory-augmented reinforcement learning strengthens LLM agents' ability to solve complex long-horizon tasks. Skills are one such form of memory, pairing instructions with an applicability condition over task types. However, retaining every skill indiscriminately as the policy improves lets obsolete or harmful entries accumulate and mislead the agent. We propose SkillForge, an agentic RL method that compiles and evolves the skill library through a fitness-driven skill lifecycle of trial, active, stable, and retired states, so that the skills and the model co-evolve throughout training. A preliminary evaluation phase first uses the base model's own rollouts to pre-retire low-fitness skills, yielding a filtered library that then seeds supervised fine-tuning. Reinforcement learning takes over from this checkpoint, and at each iteration selective retirement, stabilization, and LLM-guided mutation continue to forge the skill library alongside policy optimization. Across multiple interactive agent benchmarks, SkillForge achieves the highest aggregate success rate, delivering up to 7.8\% relative improvement over the strongest baseline while keeping the skill library compact throughout training. We introduce Skillfurnace, a dataset of 5k+ annotated records bundling retirement-filtered SFT trajectories, evolved skill libraries with fitness annotations, and retirement events with human-annotated failure categories to support research on skill quality and lifecycle management.

Reinforcement learning (RL) has been widely used to train LLM agents for multi-turn interactive tasks, but its sample efficiency is severely limited by sparse rewards and long horizons. On-policy self-distillation (OPSD) alleviates this by providing dense token-level supervision from a privileged teacher that has access to ground-truth answers. However, such fixed privileged information cannot capture the diverse valid strategies in agent tasks, and naively combining OPSD with RL often leads to training collapse. To address these limitations, we introduce Skill-SD, a framework that turns the agent's own trajectories into dynamic training-only supervision. Completed trajectories are summarized into compact natural language skills that describe successful behaviors, mistakes, and workflows. These skills serve as dynamic privileged information conditioning only the teacher, while the student always acts under the plain task prompt and learns to internalize the guidance through distillation. To stabilize the training, we derive an importance-weighted reverse-KL loss to provide gradient-correct token-level distillation, and dynamically synchronize the teacher with the improving student. Experimental results on agentic benchmarks demonstrate that Skill-SD substantially outperforms the standard RL baseline, improving both vanilla GRPO (+14.0%/+10.9% on AppWorld/Sokoban) and vanilla OPD (+42.1%/+40.6%).

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for post-training large language models on reasoning tasks. To avoid the complexity of training explicit critics, algorithms like GRPO estimate advantages directly from sparse on-policy rollouts. However, the efficacy of this approach relies heavily on the accuracy of the baseline estimator. In this work, we argue that standard methods, whether relying on prompt-local averages or shrinking toward a single global mean, are statistically mis-specified for reasoning domains. By enforcing a uniform prior, these methods ignore the latent skill structure of tasks, introducing estimation bias that conflates dissimilar problems. To address this, we propose Skill-Coupled Policy Optimization (SCPO), a structured baseline estimation framework that refines the shrinkage prior by leveraging task correlations. Instead of a generic global target, SCPO performs shrinkage toward skill-coupled group targets, thereby aligning the baseline with the local difficulty landscape. To further stabilize these targets under limited on-policy observations, SCPO incorporates a lightweight history tracker. Experiments across diverse models and tasks show that SCPO consistently improves training stability and performance over both prompt-level and global shrinkage baselines.

Reinforcement Learning (RL) algorithms often rely on data collected from a target environment, which can be expensive and noisy. Although simulators can generate large amounts of additional data, standard simulation mixing approaches—which directly combine simulated and real experience—can degrade performance due to simulator inaccuracies, especially when they bias the solution away from the optimal policy. We propose a conservative simulation-aided learning method based on control variates, where simulated data is used solely to reduce variance in the learning process rather than directly shaping the learning target. We develop our approach for Policy-Evaluation, which is a fundamental building block of RL algorithms. Then we combine it within the actor-critic framework towards policy improvement, as well as extend it towards offline RL scenarios. Experiments on multiple MuJoCo benchmark tasks demonstrate improvements over standard simulation mixing approaches, highlighting control variates as an effective tool for simulation-aided reinforcement learning.

Large language models (LLMs) can generate student-like responses fluently enough to serve as simulated students, i.e., virtual learners used to train and evaluate AI tutors and human educators. Yet such simulators are typically evaluated by output similarity to real students, not by whether they behave like students with coherent misconceptions during interaction. We introduce a controlled framework for evaluating misconception faithfulness, whether a simulator maintains a misconception-driven belief state and updates selectively when feedback addresses the underlying misconception. Central to our framework is a misconception-contrastive feedback protocol that compare targeted feedback against two controls: misaligned feedback, targeting a different but plausible misconception, and generic feedback, which only signals that the answer is incorrect. We propose Selective Flip Score (SFS), which quantifies how much more often a simulator flips its answer under targeted feedback than under the contrastive controls. Across seven LLMs (4B–120B), multiple datasets, and prompting strategies, simulators exhibit near-zero SFS, correcting their answers at similarly high rates regardless of feedback relevance to the true misconception. Further analysis reveals a sycophantic failure mode: models behave less like students with stable misconceptions and more like problem solvers who treat any corrective signal as a cue to abandon the simulated misconception and recompute from internal knowledge. To improve faithfulness, we develop a post-training pipeline spanning supervised fine-tuning (SFT), preference optimization, and reinforcement learning (RL) with an SFS-aligned reward. SFT yields notable gains up to +0.56; SFS-aligned RL provides more consistent further improvements than preference optimization. Our results establish misconception faithfulness as a challenging but trainable property of student simulators, motivating a shift from static output matching toward interaction- and belief-aware student modeling.

Recurrent spiking neural networks (RSNNs) are a promising substrate for energy-efficient control policies, but training them for high-dimensional, long-horizon reinforcement learning remains challenging. Population-based, gradient-free optimization circumvents backpropagation through non-differentiable spike dynamics by estimating gradients. However, with finite populations, high variance of these estimates can induce harmful and overly aggressive update steps. Inspired by trust-region methods in reinforcement learning that constrain policy updates in distribution space, we propose **Signal-Adaptive Trust Regions (SATR)**, a distributional update rule that constrains relative change by bounding KL divergence normalized by an estimated signal energy. SATR automatically expands the trust region under strong signals and contracts it when updates are noise-dominated. We instantiate SATR for Bernoulli connectivity distributions, which have shown strong empirical performance for RSNN optimization. Across a suite of high-dimensional continuous-control benchmarks, SATR improves stability under limited populations and reaches competitive returns against strong baselines including PPO-LSTM. In addition, to make SATR practical at scale, we introduce a bitset implementation for binary spiking and binary weights, substantially reducing wall-clock training time and enabling fast RSNN policy search.

Existing guarantees for misspecified kernelized bandit optimization pay for misspecification through kernel complexity: in generic offline bounds, the misspecification level $\varepsilon$ is multiplied by $\sqrt{d_\mathrm{eff}}$, where $d_\mathrm{eff}$ is the kernel effective dimension, while in online regret bounds, the corresponding penalty is $\sqrt{\gamma_n}\,n\varepsilon$, where $\gamma_n$ is the maximum information gain after $n$ rounds of interaction. In this work, we show that, for a large class of kernels, the misspecification amplification can be reduced to logarithmic or polylogarithmic growth. In the offline setting, we first prove high-probability simple-regret bounds whose misspecification term is governed by a spectral Lebesgue constant. This yields logarithmic amplification for one-dimensional monotone spectra and polylogarithmic amplification for multivariate Fourier-diagonal product kernels. In the online setting, we modify a domain-splitting algorithm and prove a cumulative regret bound of $\widetilde{\mathcal O}(\sqrt{\gamma_n n}+n\varepsilon)$ under mild localized eigendecay assumptions, removing the extra $\sqrt{\gamma_n}$ factor from the misspecification term. The common principle is localization: spectral localization controls the Lebesgue constant of the offline approximation operator, while domain splitting implements the spatial analogue of this mechanism in the online setting, preventing local misspecification errors from being amplified globally.

A universal controller for any robot morphology would greatly improve computational and data efficiency. Steps have been made towards such multi-robot control by utilizing *contextual* information about the properties of individual robots and exploiting their modular structure in the architecture of deep reinforcement learning agents. When the robots have highly dissimilar morphologies, however, this becomes a challenging problem, especially when the agent must generalize to new, unseen robots. In this paper, we posit that contextual features are often only partially available, but that they can be recovered through modular interactions. This can allow for better multi-robot control and generalization to contexts that are not seen during training. To this extent, we implement a transformer-based architecture with shared modular recurrence and evaluate its (generalization) performance on a large set of MuJoCo robots. The results show a substantial improvement in zero-shot generalization performance on robots with unseen dynamics, kinematics, and topologies, in four different environments.

Itemwise calibration does not guarantee calibration of the decision it feeds. Many systems output one score per item in a partially observed group, but the downstream action asks whether enough hidden items cross a threshold: show a slate, defer a panel decision, or inspect a batch. If those hidden items share an unresolved user, patient, or batch state, averaging that state before aggregation preserves means but makes the aggregate posterior too concentrated. We call this failure the serialization tax. For Bernoulli de Finetti posteriors, the exact hidden-count law dominates any independent completion with the same total predictive mean in convex order, yielding variance, total-variation, threshold, and Bayes-value gaps. The same interface loss extends to non-binary outcomes through empirical-measure Laplace functionals and to heterogeneous slates through a shared-covariance identity. TaxScore turns the theory into a routing rule: use cheap marginal interfaces when hidden-block ambiguity is far from the decision boundary, and propagate a shared posterior sample when it is near. On pre-specified, hidden-label-free MovieLens 1M/10M audits, replacing Transformer count heads with existing stochastic shared-latent heads improves high-TaxScore negative log-likelihood (NLL)/utility from $0.673/0.696$ to $0.629/0.723$ and from $0.670/0.691$ to $0.639/0.715$; routing richer heads only on flagged slates gives $0.006$-$0.012$ full-population utility gains.

Vision-Language-Action (VLA) models offer a promising paradigm for end-to-end autonomous driving by connecting visual scene understanding, linguistic reasoning, and low-level trajectory generation. However, existing driving VLAs typically use language only as static commands or supervised rationales, while treating trajectory prediction as an open-loop generation problem. As a result, they lack an explicit mechanism for understanding the consequences of their own actions, which limits robustness under closed-loop distribution shift and makes reinforcement learning inefficient in long-tail scenarios. We propose SelfCritic-VLA, a self-evaluative VLA framework that unifies trajectory generation, language-grounded action critique, and trajectory refinement. Given a driving scene, the model first generates an initial trajectory, then produces a language-grounded critique that describes the potential safety, progress, and rule-compliance consequences of executing that trajectory, and finally refines the trajectory conditioned on this critique. To train this capability, we construct a large-scale counterfactual trajectory dataset with multi-source candidate trajectories and semantic critiques, and perform hybrid supervised fine-tuning for planning, action evaluation, and refinement. Building on this initialization, we introduce a self-critic reinforcement learning procedure based on GRPO, where critique-conditioned refinements are filtered by grounded driving rewards and distilled back into the base policy. Experiments on NAVSIM and Bench2Drive show that SelfCritic-VLA consistently improves both open-loop and closed-loop driving performance over strong baselines, demonstrating the effectiveness of language-guided self-refinement for driving policy optimization.

Contrastive reinforcement learning (CRL) learns goal-conditioned Q-values through a contrastive objective over state-action and goal representations, removing the need for hand-crafted reward functions. Despite impressive success in achieving viable self-supervised learning in RL, all existing CRL algorithms rely on off-policy optimisation and are mostly constrained to continuous action spaces, with little research invested in discrete environments. This leaves CRL disconnected from widely used and effective, modern on-policy training pipelines adopted across both single-agent and multi-agent RL in continuous and discrete environments. To establish a first connection, we introduce Contrastive Proximal Policy Optimisation (CPPO). CPPO is an on-policy contrastive RL algorithm that derives policy advantages directly from contrastive Q-values and optimises them via the standard PPO objective, without requiring a reward function or a replay buffer. We evaluate CPPO across continuous and discrete, single-agent and cooperative multi-agent tasks. Whilst the existence of an on-policy approach is inherently useful, we additionally observe that CPPO not only significantly outperforms the previous CRL baselines in 14 out of the 18 tasks benchmarked, but matches or exceeds reward-based PPO's performance in 12 out of the 18 tasks without any reward signal.

Protein sequence optimization under tight oracle budgets requires methods that explore vast combinatorial spaces while making each evaluation informative. Existing reinforcement learning and off-policy generative approaches often degrade under surrogate noise, and position-agnostic mutation proposals risk disrupting functionally critical residues. We introduce SILO, a trajectory-level self-improvement imitation framework for oracle-budgeted protein design. SILO uses a hierarchical edit policy that decomposes each mutation into a position choice followed by a residue choice. In each active-learning round, the policy samples candidate trajectories via incremental stochastic beam search without replacement (SBS), and a UCB-based proxy ensemble, combined with an alanine-scan fitness score (AFS), selects candidates with functionally relevant edits for in silico oracle evaluation. The policy is then updated by next-action cross-entropy imitation on the round’s best oracle-labeled trajectories, avoiding value-function estimation. Across eight reproduced protein fitness landscapes and five strong baselines from prior work, SILO achieves the highest maximum and top-100 mean fitness on 8 of 8 landscapes within our evaluations, often exhibiting faster early-stage improvement. In low-data and noisy-proxy stress tests on two landscapes per setting, SILO remains competitive or best when several baselines degrade. Ablations show that SBS with AFS account for much of the gains, with iterative imitation providing additional improvement.

Developing self-evolving agents that learn continuously from streams of information represents an emerging frontier in artificial intelligence. This capability is critical for efficiently updating computationally expensive systems such as foundation models. When agents learn by exploring the environment and interacting with other agents to achieve multiple tasks over time, two sources of non-stationarity emerge: shifting task distributions and co-evolving peers. However, developing frameworks for continual multi-agent reinforcement learning (CMARL) and effective policy improvement remains challenging. In this paper, we argue that training self-evolving agents should incorporate two classes of model-based principles: internal models that allow an agent to compare policy checkpoints, monitor its learning capacity, and represent its current objective; and external models that learn environment dynamics and the behavior of other agents. We anticipate that discussion on these principles will facilitate the design of algorithms, architectures, and benchmarks that explicitly evaluate model-based components in CMARL.