NeurIPS 2026 论文列表

Aligning Large Language Models (LLMs) with human preferences typically relies on external supervision, which faces critical limitations: human annotations are scarce and subjective, reward models are vulnerable to reward hacking, and self-evaluation methods suffer from prompt sensitivity and biases. In this work, we propose **stable rank**, an intrinsic, annotation-free quality signal derived from model representations. Stable rank measures the effective dimensionality of hidden states by computing the ratio of total variance to dominant-direction variance, capturing quality through how information distributes across representation dimensions. Empirically, stable rank achieves 84.04\% accuracy on RewardBench and improves task accuracy by an average of 14.1\% relative to greedy decoding via Best-of-N sampling. Leveraging this insight, we introduce **Stable Rank Group Relative Policy Optimization (SR-GRPO)**, which uses stable rank as a reward signal for reinforcement learning. Without external supervision, SR-GRPO improves Qwen2.5-1.5B-Instruct by 10\% on STEM and 19\% on mathematical reasoning, outperforming both learned reward models and self-evaluation baselines. Our findings demonstrate that quality signals can be extracted from internal model geometry, offering a path toward scalable alignment without external supervision.

Logical reasoning remains a major challenge for large language models (LLMs), particularly on structured problems that require precise constraint tracking, consistency preservation, and multi-step deduction. This challenge is especially acute for small-scale LLMs, which are more prone to producing inconsistent, redundant, or brittle reasoning trajectories. Existing approaches for improving logical reasoning largely optimize for final-answer correctness, providing only weak supervision over the intermediate reasoning process. In this work, we propose SPRING, a solver-guided reinforcement learning framework for logical reasoning that uses an SMT solver as a training-time verifier of intermediate reasoning steps to provide process-level supervision. SPRING introduces the notion of a novel reasoning step, namely, a step that is logically valid, consistent with the evolving reasoning state, and not already implied by previously accepted non-contradictory deductions. Based on this solver-based assessment, we design process rewards that encourage novel inferential progress while penalizing contradictory and uninformative reasoning steps. Evaluation results on two logical reasoning benchmarks, ZebraLogic and AR-LSAT, show that SPRING consistently outperforms baseline LLMs and outcome-only reward baselines. On ZebraLogic, SPRING improves puzzle accuracy by up to 49.71 and 15.43 points over the base LLM and outcome-only reward baseline, respectively. On AR-LSAT, SPRING improves the overall average score by up to 64.93 and 12.14 points over the base LLM and best-performing outcome-only reward baseline, respectively.

Scaling test-time compute has driven the recent advances in the reasoning capabilities of large language models (LLMs). However, increased compute often comes at the expense of higher user-facing latency, directly impacting user experience. Current test-time scaling methods primarily optimize for accuracy based on total compute resources (FLOPs), often overlooking latency constraints. To address this gap, we propose SPECS, a latency-aware test-time scaling method. SPECS uses a smaller, faster model to generate multiple candidates for each step of reasoning in parallel, and evaluates these candidates using a larger model and a quality critic. We design a theoretically grounded soft verification strategy to select a high-quality candidate to continue the generation. SPECS also employs a dynamic switching mechanism to use speculative drafts only for easier steps to maintain reasoning accuracy. Empirical results on a diverse and challenging set of reasoning and alignment benchmarks show that SPECS matches or surpasses the accuracy of SOTA test-time scaling methods while reducing latency by up to ~26%. Our theoretical analysis shows that as the amount of parallel compute scales, SPECS converges to an optimum of a KL-regularized reinforcement learning problem, a common objective for aligning LLM generation given a reward signal.

Vision-language models (VLMs) have shown impressive capabilities across diverse tasks, motivating efforts to leverage these models to supervise robot learning. However, when used as evaluators in reinforcement learning (RL), today’s strongest models often fail under partial observability and distribution shift, enabling policies to exploit perceptual errors rather than solve the task. We introduce SOLE-R1 (Self-Observing LEarner), a video-language reasoning model explicitly designed to serve as the sole reward signal for online RL. Given only raw video observations and a natural-language goal, SOLE-R1 performs per-timestep spatiotemporal chain-of-thought (CoT) reasoning and produces dense estimates of task progress that can be used directly as rewards. To train SOLE-R1, we develop a large-scale video trajectory and reasoning synthesis pipeline that generates temporally grounded CoT traces aligned with continuous progress supervision. This data is combined with foundational spatial and multi-frame temporal reasoning, and used to train the model with a hybrid framework that couples supervised fine-tuning with RL from verifiable rewards. Across four different simulation environments and a real-robot setting, SOLE-R1 enables zero-shot online RL from random initialization: robots learn previously unseen manipulation tasks without ground-truth rewards, success indicators, demonstrations, or task-specific tuning. SOLE-R1 succeeds on 24 unseen tasks and substantially outperforms strong vision-language rewarders, including GPT-5 and Gemini-3-Pro, while exhibiting markedly greater robustness to reward hacking.

Classical ReLU-based Input Convex Neural Networks (ICNNs) are equivalent to the optimal value functions of Linear Programming (LP). This intrinsic structural equivalence restricts their representational capacity to piecewise-linear polyhedral functions. To overcome this representational bottleneck, we propose the SOC-ICNN, an architecture that generalizes the underlying optimization class from LP to Second-Order Cone Programming (SOCP). By explicitly injecting positive semi-definite curvature and Euclidean norm-based conic primitives, our formulation introduces native smooth curvature into the representation while preserving a rigorous optimization-theoretic interpretation. We formally prove that SOC-ICNNs strictly expand the representational space of ReLU-ICNNs without increasing the asymptotic order of forward-pass complexity. Extensive experiments demonstrate that SOC-ICNN substantially improves function approximation, while delivering competitive downstream decision quality. The code is available at \url{https://anonymous.4open.science/r/SOC-ICNN-4B18/}.

Model-based reinforcement learning (MBRL) is sample-efficient but struggles in sparse reward settings. A critical bottleneck arises from the lack of informative gradients in sparse settings, where standard reward models often yield flat landscapes that struggle to guide planning. To address this challenge, we propose **S**haping **L**andscapes with **O**ptimistic **P**otential **E**stimates (**SLOPE**), a novel framework that shifts reward modeling from predicting sparse scalars to constructing informative potential landscapes. SLOPE employs optimistic distributional regression to estimate high-confidence upper bounds, which amplifies rare success signals and ensures sufficient exploration gradients. Evaluations on **30+** tasks across **5** benchmarks and **real-world robotic deployments**, demonstrate that SLOPE consistently outperforms leading baselines in **fully sparse**, **semi-sparse**, and **dense** rewards. The project code is available at https://anonymous.4open.science/r/SLOPE-CF13.

Reinforcement learning from verifiable rewards (RLVR) trains reasoning language models with group-relative objectives such as GRPO. However, it leaves two sources of signal underused: every token in a rollout receives the same response-level advantage, and groups in which all rollouts succeed or all fail have zero group-relative advantage and are typically discarded. We observe that same prompt rollouts are not independent reward labels but sibling attempts: they share scaffolding, diverge at uncertain decisions, and sometimes reconverge. In this paper, we propose Sibling-Relative Advantage Shaping (SIRAS), a lightweight advantage-shaping method that exploits this structure. SIRAS segments rollouts into chunks, aligns sibling chunks with banded dynamic time warping, and reshapes the flat GRPO advantage across tokens via soft-divergence weights that upweight chunks distinguishing a rollout from its siblings and downweight shared ones. This reshaping preserves each rollout’s token-averaged GRPO advantage exactly and reduces to GRPO when weights are constant. The same sibling structure produces calibrated residuals for all-correct and all-wrong groups, reclaiming training signal that group-relative baselines discard. Across six reasoning benchmarks, SIRAS yields the largest gains on AIME-style competition math, including +12.2 on AIME25 with Qwen3-8B and +23.3 on AIME26 in our backbone-scaling study, without extra rollouts, process reward models, or step annotations.

Recent mixed-policy optimization methods for LLM reasoning that interleave or blend supervised and reinforcement learning signals report improvements over the standard SFT-then-RL pipeline. We show that numerous recently published research papers rely on a faulty baseline caused by two distinct bugs: a CPU-offloaded optimizer bug in DeepSpeed that silently drops intermediate micro-batches during gradient accumulation (affecting multiple downstream frameworks including TRL, OpenRLHF and Llama-Factory), and a loss aggregation bug in OpenRLHF that incorrectly weights per-mini-batch losses. Together they suppress SFT performance, with the optimizer bug accounting for most of the gap and the loss aggregation bug contributing a smaller additional effect. Once corrected, the standard SFT-then-RL pipeline surpasses every published mixed-policy method we evaluate by +3.8 points on math benchmarks with Qwen2.5-Math-7B and by +22.2 points with Llama-3.1-8B. Even a truncated variant with just 50 RL steps outperforms mixed-policy methods on math benchmarks while using fewer FLOPs.

Large language models (LLMs) are rapidly shifting toward agents that solve tasks through diverse interfaces, including web and graphical user interfaces (GUIs). Among these, the terminal command line provides a text-based, general-purpose interface, covering tasks from system operations to data science and machine learning. However, scaling terminal-agent training remains challenging, as it requires diverse and coherent task instructions, executable environments, and reliable verification, while lacking naturally grounded supervision data. In this work, we propose SETA, a scalable framework for generating verifiable terminal environments for reinforcement learning (RL). The framework consists of two pipelines sharing a unified verification mechanism: SETA-Synth converts diverse sources into standardized RL environments, and SETA-Evol further expands from existing environments with adaptive control of difficulty and diversity. Together, we construct and release SETA-Env, the largest open-source verifiable terminal RL dataset to date, containing over $4{,}500$ environments. We evaluate our dataset by training a Qwen3-8B model with GRPO on SETA-Env, achieving $12$\% pass rate on Terminal-Bench 2.0, the best reported result for an RL-trained model at the 8B scale. The results demonstrate that SETA-Env provides high-quality training environments for terminal agents and serves as a valuable resource for advancing research on terminal-based agent learning.

Action chunking improves exploration and accelerates value propagation in long-horizon reinforcement learning, but naively applying off-policy methods to the temporally extended action space at reduced decision frequency offsets these gains, leading to poor sample efficiency. Existing action chunking methods address these issues through computationally expensive critic-only approaches or by relying on offline data. We introduce SEAR, a sample-efficient off-policy algorithm that enables online reinforcement learning with action chunks by addressing both challenges. To handle high-dimensional action sequences, SEAR employs a causal transformer critic trained with multi-horizon targets that provide a training signal for every prefix of an action chunk, effectively increasing the useful gradients per sample. To restore decision frequency, SEAR operates with a receding horizon and random replanning, ensuring uniform state coverage while combining the fast value propagation of large chunks with the reactivity of small ones. SEAR outperforms state-of-the-art online reinforcement learning methods including SimbaV2 on challenging Metaworld manipulation tasks. Beyond online RL, applying SEAR to the offline-to-online method QC improves its performance on the demanding OGBench cube-triple tasks.

Search-augmented reasoning agents interleave internal reasoning with calls to an external retriever, and their performance relies on the quality of each issued query. However, under outcome-reward reinforcement learning, every search decision in a rollout shares the same trajectory-level reward, leaving individual queries without step-specific credit. Recent process-supervision approaches address this gap by drawing step-level signals from outside the policy, relying either on a much larger teacher model, or on sub-question annotations produced by a stronger external system. In contrast, we propose \textbf{SD-Search}, which derives step-level supervision \emph{from the policy itself} through hindsight self-distillation, requiring neither an external teacher nor additional annotations. In SD-Search, a single model plays two roles that differ only in conditioning: a student that sees only the context available at inference time, and a teacher that additionally conditions on a compact hindsight block summarizing the search queries and final outcomes of a group of rollouts sampled from the same question. Since teacher knows how each rollout unfolded and which ones succeeded, its query distribution implicitly marks which decisions were worth making, and the student is trained to recover this behavior by minimizing the token-level Jensen--Shannon divergence to the teacher at search-query positions. This layers a dense, step-level signal on top of GRPO's coarse trajectory reward. Crucially, this signal requires only a single teacher forward pass per rollout at training time, all absorbed within the standard RL loop without external model inference, auxiliary annotation pipeline, or additional training. On seven single-hop and multi-hop QA benchmarks, SD-Search reaches $0.428$ average Exact Match with Qwen2.5-3B, matching the leading process-supervision baseline without its external teacher, and $0.476$ with Qwen2.5-7B, surpassing all outcome-reward and process-supervision baselines on the average.

The alignment of Large Language Models (LLMs) via reinforcement learning from human feedback (RLHF) typically relies on Proximal Policy Optimization (PPO), which uses a uniform clipping hyperparameter $\epsilon$ to maintain a stable trust region. However, we demonstrate that this uniform constraint is ill-suited for the non-uniform semantic topology of natural language. In ``semantically dense'' regions, where candidate tokens act as semantic synonyms within the current context, noisy advantage estimates can destabilize the fixed trust region, leading to *policy collapse*, a severe loss of generative diversity and policy entropy. To resolve this, we propose *Structure-Aware Mirror Proximal Policy Optimization (SAMPPO)*, which introduces an adaptive trust region governed by local semantic density. SAMPPO generalizes the trust-region framework to non-uniform semantic spaces by adapting the mirror-descent geometry to the local action topology. SAMPPO uses a per-timestep *Semantic Isolation Score* to dynamically tighten the clipping range in dense regions and relax it in sparse ones. We provide theoretical foundations for this approach, deriving a *Structure-Aware Monotonic Improvement* bound and establishing a formal link to structure-aware margin-shifted $H$-consistency theory. Empirically, SAMPPO prevents the entropy collapse observed in standard PPO baselines and significantly improves alignment. Evaluated on the Anthropic HH-RLHF and NVIDIA HelpSteer benchmarks across multiple model architectures (Llama-3-8B and Gemma-2-2B), SAMPPO achieves robust, zero-shot generalizable alignment gains. Crucially, it strictly outperforms standard PPO and recent empirical adaptive heuristics (BAPO), demonstrating competitive performance against direct preference baselines including DPO and SimPO on the alignment-diversity Pareto frontier.

Multi-stage ML inference pipelines are difficult to autoscale because of heterogeneous resources, cross-stage coupling, and dynamic bottleneck migration. Production controllers ignore cross-stage effects, and learning-based autoscalers require offline training data that is unavailable for new pipelines. We present SAIR, an autoscaler for the zero-shot deployment regime that uses an LLM as an in-context reinforcement learning controller, improving its policy online from reward-labeled experience tuples in the LLM's context with no gradient updates. SAIR combines Pareto-dominance reward shaping with a provable separation margin between dominated and non-dominated actions, Bayesian-surprise experience retrieval, and a user-space CUDA interception layer that delivers continuous GPU rate control with sub-millisecond actuation overhead. We prove a regret bound that decomposes into retrieval coverage and LLM selection components, each mapping to a concrete SAIR mechanism. On four ML serving pipelines under three workload patterns, SAIR achieves the best or tied-best P99 latency in every configuration (up to 50\% reduction over the next baseline); billable cost is reduced through replica consolidation, and effective cost (which captures GPU rate-control savings under co-location or MIG) by up to 97\%. SAIR reaches 86\% bottleneck detection accuracy without any offline training.

Reinforcement learning from human feedback (RLHF) with preference-based reward models often exhibits unstable training dynamics. A key contributing factor is that standard RLHF relies on a single sequence-level scalar reward, which is propagated to token-level policy updates and leaves credit assignment within a response inherently ambiguous. Recent work has attempted to address this issue by refining rewards into denser token-level supervision, often relying on the implicit assumption that finer-grained credit assignment improves optimization. We argue that this assumption is incomplete: when preference signals are noisy and only defined at the response level, overly fine-grained reward refinement can amplify reward uncertainty and destabilize learning. To address this problem, we propose a granularity-aware principle for hierarchical credit assignment, emphasizing stability-oriented reward design rather than maximal allocation precision. Under this principle, sentences serve as a natural intermediate granularity, balancing semantic coherence with robustness to token-level noise. Guided by this view, we introduce S2T-RLHF. This sentence-to-token reward decomposition framework first allocates sequence-level preference rewards across sentences and then applies bounded token-level refinement within each sentence, without reward-model retraining or token-level supervision. Experiments across multiple datasets and optimization settings show that S2T-RLHF improves training stability and robustness while maintaining competitive preference alignment.

While dataset distillation has witnessed significant advances in image classification, its extension to dense prediction tasks such as semantic segmentation is plagued by two core bottlenecks: bi-level optimization-based approaches suffer from prohibitive computational costs stemming from pixel-wise gradient unrolling, whereas proxy-based generative methods merely infuse visual textures into fixed semantic masks derived from real data, thus failing to condense rich semantic knowledge into a compact set of informative, novel scene generations. To address both bottlenecks simultaneously, we propose a novel Sample-Set Reinforcement Learning (S$^2$-RL) framework for generative semantic segmentation dataset distillation (SSDD). Specifically, S$^2$-RL formulates SSDD as a diffusion model based text-to-image generation paradigm, enabling flexible generation of images with informative, novel semantic distributions via a concise text prompt. Furthermore, we fine-tune a diffusion policy via GRPO, leveraging a sample-set dual reward paradigm: for individual samples, it enforces semantic alignment with input text prompts and intra-class feature diversity; for sample sets, it maximizes inter-sample semantic distribution diversity by solving a contextual multi-armed bandit problem. These design choices enable S$^2$-RL to distill large-scale semantic segmentation datasets into a compact set of generated samples that encapsulate diverse, informative semantic knowledge without the need of pixel-wise gradient unrolling. Extensive experiments on the ADE20K and COCO datasets demonstrate that S$^2$-RL achieves substantial and consistent improvements over state-of-the-art baselines, establishing a new benchmark for SSDD.

Recent progress in large language models (LLMs) has led to impressive performance across diverse domains. However, their deployment in critical areas such as healthcare, law, and education raises serious concerns regarding the potential generation of harmful content. While reinforcement learning (RL)-based safety alignment methods have been extensively studied, they face persistent challenges, including sparse reward signals, limited interpretability, and limited transferability across domains. In this paper, we introduce Rubric‑Align, a framework that replaces static binary scalar rewards with dynamic, natural‑language evaluation rubrics. Instead of assigning a single reward to each response, Rubric‑Align provides fine‑grained and interpretable feedback, thereby alleviating reward sparsity and improving the transparency of the alignment signal. A key feature of Rubric-Align is its dynamic rubric evolution mechanism. Rather than using fixed reward templates, Rubric-Align periodically refines prompt-specific rubrics based on the current policy behavior, keeping the supervision signal informative and aligned with the model’s evolving failure modes. Experiments on safety benchmarks and vertical-domain settings show that Rubric-Align improves robustness against harmful and jailbreak prompts while largely preserving general capabilities, suggesting that natural-language rubrics provide a transferable interface for adapting safety supervision to new safety subdomains.

Diffusion models have emerged as a powerful tool across diverse domains. However, their purely data-driven nature can produce samples that violate domain-governing constraints. We introduce a plug-and-play Reinforcement Learning framework that optimizes initial noise samples in the latent space of frozen, pre-trained diffusion models. Leveraging the near-spherical geometry of high-dimensional Gaussian distributions, we introduce a novel rotation-matrix-based scheme for efficient latent space exploration. This steers the model toward more feature-preserving outputs, guided by task-specific rewards. We evaluate our method on three diffusion models: one trained on solutions of the Darcy Flow PDE, one on a synthetic dataset with complex structural features, and a text-conditioned one. Across all three settings, our framework yields significant improvements in sample quality, achieving a ${\sim}25\\%$ relative reduction in PDE residual, up to a ${\sim}44\\%$ relative improvement on the synthetic dataset's feature-alignment metric, and up to a ${\sim}80\\%$ relative improvement on human preference, compared to the vanilla diffusion models. Finally, we show that rotation-matrix-based exploration significantly outperforms unconstrained exploration, validating our geometry-aware approach and establishing a more effective method for latent space control.

Modern data analysis pipelines increasingly rely on large datasets assembled through retrieval, scraping, automated filtering, and post-hoc curation, where the observed sample is often a contaminated mixture rather than the target population. Standard distributional objectives either ignore contamination and converge to the wrong target, or retain only a small trusted subset and discard most of the data. We study supervised and unsupervised distributional learning when a clean reference sample is compared with a pooled sample from $r=(1-\epsilon)q+\epsilon z$, where $q$ is the latent target distribution and $z$ is an unknown contaminant. Motivated by settings with noisy membership scores for all observations and exact labels for a small audited subset, we propose contamination-corrected maximum mean discrepancy (CC–MMD), a design-based estimator that combines proxy scores with audited residual corrections to recover the oracle MMD that would have been computed from the latent clean sample. CC–MMD requires no calibration assumption on the proxy scores and no structural assumptions on the contaminant beyond standard kernel regularity. We prove finite-population design-unbiasedness of the corrected kernel sums, establish asymptotic normality of the audit correction, and show consistency for the target discrepancy $\mathrm{MMD}^2(p,q;k)$. Empirically, CC–MMD tracks the oracle under increasing contamination, improves parameter recovery and prediction in contaminated regression, and enables generative models trained on heavily corrupted mixtures to recover the clean target distribution. These results show that contaminated data need not be discarded or treated as the estimand. With sparse trusted labels and abundant noisy supervision, distributional objectives can be made robust by design.

Data Streams are characterised as a potentially infinite source of data subject to non-stationarity. This non-stationarity, also called Concept Drift, often results in deployed predictive models having to be perpetually retrained as the underlying data generating process of the stream changes over time. Drift Detectors are a class of methods used to identify these changes, triggering these retraining procedures. Historically, Stream Learning has operated under the assumption that instances are only seen once, and cannot be stored for later use. Contemporary work relaxes this assumption yet drift detection and simple resetting procedures are still ubiquitous. Building on related work which has shown that traditional Batch Learning algorithms often outperform Stream Learning algorithms when instances can be stored, we introduce Time Neutralising Trees (TNT), a Decision Tree architecture that enables robust stream classification in settings with Concept Drift. During training, TNT will filter out ("neutralise") old training samples that are no longer relevant given the current context of the stream. If those instances become relevant again, TNT will then re-include them. We evaluate TNT, and related algorithms on 24 real and semi-real classification data streams from the USP DS repository. Results show strong evidence that TNT achieves state-of-the-art performance on a wide-range of data streams with varying concept drift types.