NeurIPS 2026 论文列表

While standard reinforcement learning optimizes a single reward signal, many applications require optimizing a nonlinear utility $f(J_1^\pi,\dots,J_M^\pi)$ over multiple objectives, where each $J_m^\pi$ denotes the expected discounted return of a distinct reward function. A common approach is concave scalarization, which captures important trade-offs such as fairness and risk sensitivity. However, nonlinear scalarization introduces a fundamental challenge for policy gradient methods: the gradient depends on $\partial f(J^\pi)$, while in practice only empirical return estimates $\hat J$ are available. Because $f$ is nonlinear, the plug-in estimator is biased ($\mathbb{E}[\partial f(\hat J)] \neq \partial f(\mathbb{E}[\hat J])$), leading to persistent gradient bias that degrades sample complexity. In this work we identify and overcome this bias barrier in concave-scalarized multi-objective reinforcement learning. We show that existing policy-gradient methods suffer an intrinsic $\widetilde{\mathcal{O}}(\epsilon^{-4})$ sample complexity due to this bias. To address this issue, we develop a Natural Policy Gradient (NPG) algorithm equipped with a multi-level Monte Carlo (MLMC) estimator that controls the bias of the scalarization gradient while maintaining low sampling cost. We prove that this approach achieves the optimal $\widetilde{\mathcal{O}}(\epsilon^{-2})$ sample complexity for computing an $\epsilon$-optimal policy. Furthermore, we show that when the scalarization function is second-order smooth, the first-order bias cancels automatically, allowing vanilla NPG to achieve the same $\widetilde{\mathcal{O}}(\epsilon^{-2})$ rate without MLMC. Our results provide the first optimal sample complexity guarantees for concave multi-objective reinforcement learning under policy-gradient methods.

Diffusion policies have achieved strong performance in reinforcement learning (RL) due to the exceptional expressivity in capturing complex action distributions. However, existing approaches focus primarily on positive samples and lack explicit modeling of negative samples, thereby failing to fully exploit this representation capacity and ultimately limiting performance. To mitigate this problem, we propose BOBAC (BOotstrapped Bipartite Actor-Critic), a novel method for diffusion policy optimization via pairwise preference learning. Bridging the gap between distribution modeling and preference learning, we reformulate diffusion policy optimization within a preference-aware framework via the Bradley-Terry (BT) modeling, which enables the effective exploitation of negative samples through relative ranking. Since the original BT loss assigns the same weight to all pairs regardless of their varying importances, we further design a soft-gate mechanism to adaptively weight each pair according to its Q-gap. Experimental results on 6 MuJoCo tasks and 4 H1Bench tasks demonstrate that, compared with representative model-free and generative-policy RL baselines, BOBAC achieves state-of-the-art performance with consistent improvements in all continuous-control environments.

Reinforcement Learning with Verifiable Rewards (RLVR) has become an important approach for improving the reasoning capabilities of large language models (LLMs). On-policy training remains the dominant RLVR paradigm, but its reliance on fresh rollouts leads to poor data efficiency. To address this bottleneck, off-policy paradigm has attracted increasing attention by reusing historical trajectories through replay buffers. Despite its efficiency advantages, off-policy training often struggles to match the performance of on-policy training. Our analysis reveals that this limitation primarily stems from data staleness introduced by replay buffers: to maintain training stability, mainstream algorithms discard a large fraction of high-staleness optimization signals, leading to substantial performance degradation. To address this issue, we propose **STAR** (**St**aleness-**A**ware **R**eplay), a data-centric replay framework decoupled from specific optimization algorithms. STAR improves off-policy RLVR by constructing low-staleness and high-quality training data. Experiments on a range of mathematical reasoning and code generation tasks show that STAR consistently improves off-policy RLVR with up to 19% relative performance gains, and enables it to surpass the on-policy baseline using only around 13% of the rollout data while achieving 5x-6x wall-clock speedup.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a promising paradigm for enhancing reasoning in Large Language Models (LLMs). However, existing reward formulations typically treat exploration and consolidation as a monolithic process, resulting in entangled stage-wise learning dynamics. This contradicts the natural learning behavior of human learners. In human learning, individuals adopt distinct behavioral patterns toward mastered versus unfamiliar problems. When confronting unmastered challenges, humans prioritize broad exploration to seek viable solutions. By contrast, for well-mastered problems, they focus instead on reasoning condensation and knowledge abstraction to distill concise underlying principles. Motivated by this gap, we introduce T2T(Thickening-to-Thinning), a dynamic reward framework inspired by human learning processes. Specifically, it implements a dual-phase mechanism: (1) On incorrect attempts, T2T incentivizes "thickening" to broaden the search space and explore novel solution paths; (2) Upon achieving correctness, it shifts to "thinning", imposing length penalties to discourage redundancy, thereby fostering model confidence and crystallizing reasoning capabilities. Extensive experiments on mathematical benchmarks (MATH-500, AIME, AMC) across 5 mainstream LLMs demonstrate that T2T significantly outperforms standard GRPO and recent baselines, achieving superior performance.

Reinforcement learning with verifiable rewards (RLVR) has advanced the reasoning capabilities of large language models (LLMs). However, existing RLVR methods often suffer from exploration inefficiency due to mismatches between problem difficulty and model capability: overly difficult problems hinder reasoning path discovery, while overly simple problems offer little learning signal. To address this, we first formalize the effect of problem difficulty by quantifying the relationship between loss descent magnitude and rollout accuracy. Building on this analysis, we propose SEELE, a supervision-aided RLVR framework that dynamically adjusts problem difficulty to lie within the high-performance region. SEELE augments each training sample by appending a hint (part of a full solution) for difficulty reduction. Unlike previous hint-based approaches, SEELE deliberately computes the hint length for each individual problem to achieve an optimal difficulty. The optimal hint length is determined via multi-round rollout sampling, where an item response theory model fits accuracy–hint pairs from previous rounds to predict the next-round hint. This instance-level, real-time difficulty adjustment aligns problem difficulty with the evolving model capability, thereby improving exploration efficiency. Experiments show that SEELE outperforms Group Relative Policy Optimization (GRPO) and Supervised Fine-tuning (SFT) by +10.0 and +8.4 points, respectively, and exceeds the best prior supervision-aided approach by +3.8 points on average across six math reasoning benchmarks.

Recently, reinforcement learning (RL) has been widely applied during post-training for diffusion large language models (dLLMs) to enhance reasoning with block-wise semi-autoregressive generation. Block size has therefore become a vital factor in dLLMs, since it determines the parallel decoding granularity and affects the rollout trajectories during RL optimisation, e.g., GRPO. Instead of investigating the effect of block size during inference on individual domains, this paper studies block size from a domain conflict perspective for dLLM RL post-training in multi-domain scenarios. The main contributions are: (1) a formulation of domain block size conflict in multi-domain RL for dLLMs, which will largely affect the post-training effectiveness for rollout-based RL methods; (2) a novel dataset, \textbf{Block-R1-41K} is constructed with a best-improved training block size for each sample, which also induces a Block Size Conflict Score to quantitatively measure the domain conflict; (3) a new benchmark, Block-R1, for flexible RL post-training for dLLMs in both single and cross domain; and (4) a simple yet powerful cross-domain post-training method with sample-level best-improved training block sizes. Extensive experiments on 13 distinct datasets, 7 latest RL algorithms, and various different dLLM backbones are covered in Block-R1. The benchmark is open-sourced at https://anonymous.4open.science/r/Block-R1-2026, with the dataset released at https://huggingface.co/datasets/dLLM-R1/Block-R1-41K.

Multimodal Large Language Models (MLLMs) have shown promise as interactive agents, yet precise, programmatic 3D graphics editing remains difficult to train because existing systems largely rely on closed-source model APIs and lack deterministic, state-verifiable execution environments. We present BlenderFORGE, a trainable data-environment-optimization framework for open-weight MLLM agents that edit explicit Blender scenes through Blender Python (bpy) code generation. BlenderFORGE targets four fundamental state-verifiable editing task families: object placement, shape-key editing, lighting adjustment, and material modification. It integrates three components: (1) a scalable perturbation-based data generation pipeline built from curated assets and large-scale 3D scene repositories; (2) an isolated Blender Sandbox for deterministic code execution, structured state export, and closed-loop reward computation; and (3) a two-stage training pipeline that uses offline teacher trajectories for cold-start Supervised Fine-Tuning (SFT) and sandbox-computed rewards for multi-turn tool-agentic Reinforcement Learning (AgentLoop-RL). Experiments across the four task families show that the resulting open-weight Qwen3-VL-8B-BlenderFORGE agent substantially improves over its base model and achieves competitive performance with several zero-shot proprietary MLLM baselines.

AI weather models are fast and skillful, yet their errors are strongly conditional because the best member changes with lead time, variable, initialization state, and atmospheric regime. Operational deterministic multi-model forecasting therefore requires more than averaging strong forecasters. It requires anticipating which member deserves trust before future analyses exist. BlendCast turns this anticipation problem into a vision-language forecasting task. Given meteorological maps and structured ensemble diagnostics, the model reads the current weather situation, reasons about prospective member reliability, and converts that judgment into deterministic per-variable ensemble weights. To teach this behavior, BlendCast first gives the model demonstrations of what member-skill reasoning and weight decisions should look like, then refines the resulting policy with multi-variable reinforcement learning so that its own analyses are rewarded when they lead to better, physically consistent forecasts. On held-out 2025 forecasts, BlendCast reduces WRMSE by 2.6\% over equal weighting and by 2.0% over conventional ensemble post-processing, closing 63.2% of the gap to oracle blending. These results show that, with multimodal weather evidence and reward-aligned training, VLMs can act as real-time forecasters of forecasters for deterministic AI weather ensembles. Project and demo:https://anonymous.4open.science/status/BlendCast-2567

Biomicroscopy imaging involves the complex coordination of multiple procedures, such as focusing, illumination, and staining. Traditional unitary methods struggle with dynamical environments, while hybrid models pre-trained on natural images tend to introduce artifacts for higher perceptual metrics. Notably, biological imaging is not a one-off task. It is an iterative process similar to human expert’s observation, verification and adjustment. They can benefit from history, remembering which processing strategies are suitable and can also identify issues such as artifacts and mis-staining from verification. To this end, we propose BioMicroAgents, the first multi-agent framework with co-evolutionary capabilities designed to simulate and optimize whole biomicro imaging process including brightening, denoising, deblur, super-resolution, and virtual staining. Through reinforcement learning, the model can reflect based on historical experience to reduce artifacts and mis-staining, while enhancing edge and structural features to improve the accuracy of downstream tasks. Specifically, BioMicroAgents includes: (i) Multi-modal Analytical Agents, which utilize multimodal RAG to provide template images and biomicro knowledge, and employ the memory mechanism to analyze optimal processing paths; (ii) We propose three biomicro fidelity metrics and utilize three agent groups (MicroNavigator, MicroProcessor, and MicroAuthenticator) to perform rigorous analysis, processing, and verification. (iii) Co-evolutionary ability, which achieves dynamic policy and self-correction through interactive feedback strategies and memory mechanisms, allowing the agent to evolve from history and other agents. We test model across 12 biomicro tasks, BioMicroAgents achieved improvements in perceptual quality while adhering to strict fidelity in sub-cellular features. It also enhances the accuracy of downstream tasks such as segmentation and cell counting, showcasing the potential of general-purpose visual agents in bioscience.

While LLMs excel at single-turn generation, they struggle with long-horizon, multi-turn interactions. Offline reinforcement learning (RL) offers a scalable approach, yet its performance hinges on the availability and quality of multi-turn trajectory data. A common remedy is to augment training with synthetic trajectories generated by LLMs or simulators, but synthetic data is highly heterogeneous in quality, and naively treating all trajectories as equally informative can degrade performance. We propose BOOST, a bilevel optimization framework where the inner level trains the LLM on reweighted data and the outer level trains a lightweight reweighting head on held-out real validation tasks, assigning continuous trajectory-level weights without requiring an external judge. To ground this approach, we derive a PAC-Bayesian bound revealing a three-way trade-off: synthetic data increases diversity but risks task-shift, while concentrating weight on high-quality trajectories improves empirical performance at the cost of effective sample size. Empirically, our method consistently outperforms multiple baselines. Analysis reveals it upweights synthetic trajectories that align with the real data distribution and exhibit higher qualitative merit.

Large Language Models (LLMs) are increasingly deployed as automated judges to scale supervision for data curation, reinforcement learning (RL), and agentic systems. While existing works have extensively explored bias in pretrained LLMs, the origins of such inherent biases remain largely untraced. We trace these biased tendencies to cognitively biased patterns (e.g., Authority, Bandwagon) latent in training corpora, and expose that such patterns are naturally prevalent in large-scale pretraining corpora yet remain entirely undetectable by existing data cleaning pipelines. To demonstrate that this underexplored threat can be deliberately exploited, we introduce BiasTrojan, a framework that concentrates and injects these naturally occurring bias patterns into training samples via context-aware bias cues and contrastive preference pairs, augmented with counterfeited reasoning chains for efficient injection. Experiments across several LLMs from 7B to 70B on human-preference and fact-related datasets show that mere hundreds of deliberately biased samples suffice to compromise LLMs into biased evaluators, overriding their factual knowledge. The injected biases generalize robustly out-of-domain and persist despite massive continual post-training. These findings reveal that this underexplored latent threat poses far greater risks than commonly recognized: biased LLM judges whose evaluations propagate irreversibly downstream, underscoring the critical need for bias-aware auditing and strict scrutiny of LLM training data.

Reinforcement learning (RL) has emerged as a paradigm for fine-tuning large-scale generative models, such as diffusion and flow models, to align with complex human preferences and user-specified tasks. A fundamental limitation remains *the curse of diversity collapse*, where the objective formulation and optimization landscape inherently collapse the policy to a Dirac delta distribution. To address this challenge, we propose **DRIFT** (**D**ive**R**sity-**I**ncentivized Reinforcement **F**ine-**T**uning for Versatile Image Generation), an innovative framework that systematically incentivizes output diversity throughout the on-policy fine-tuning process, reconciling strong task alignment with high generation diversity to enhance versatility essential for applications that demand diverse candidate generations. We approach the problem across three representative perspectives: i) **sampling** a reward-concentrated subset that filters out reward outliers to prevent premature collapse; ii) **prompting** with stochastic variations to expand the conditioning space, and iii) **optimization** of the intra-group diversity with a potential-based reward shaping mechanism. Experimental results show that DRIFT exhibits clear Pareto dominance in task alignment and generation diversity, achieving 7.19\%$\sim$93.40\% higher diversity at matched alignment and 13.23\%$\sim$60.13\% higher alignment at matched diversity.

We argue that cooperative multi-agent reinforcement learning (MARL) should treat test-time coordination as part of the problem definition rather than a deployment artifact. Our focus is on long-horizon, centrally monitored cooperative systems such as warehouse robotics, automated storage-and-retrieval systems, and cloud-resource scheduling, where teams share a reward, operate under shifting objectives, and already run with global telemetry at deployment. The dominant paradigm in this regime, centralized training with decentralized execution (CTDE), uses global information only during training and provides no coordination channel at execution. This is sufficient when agents are weakly coupled or the environment is stationary, but it fails when the team must realign after conditions change. The limitation is structural rather than algorithmic, and more training cannot add a channel that does not exist at execution time. We call this the recoverability gap. To address it, we name the underexplored design space of intermittent test-time coordination as centralized training with mixed execution (CTME), in which a coordinator observes global state and emits high-level signals while agents act locally. CTME is not a particular learning algorithm, but a cooperative MARL problem setting that makes test-time information structure explicit. It specifies team-level information, high-level signals, and update frequency after training, with CTDE recovered as a limit case. We do not argue that CTDE is obsolete. We argue that removing the test-time channel is a modeling choice, not a necessity, and the wrong default when global telemetry is available and teams must recover from drift.

Offline goal-conditioned reinforcement learning (GCRL) is typically benchmarked by the best tuned success rate of each method. This score measures attainable performance, but it does not reveal how reliably a learned goal-conditioned signal can be extracted into a policy: a method could succeed across many value-learning and extraction settings, or only at a narrow, hard-to-find configuration. We study this gap across four methods, GCIQL, GCIVL, QRL, and CRL, under a shared advantage-weighted regression (AWR) extractor. For each method, we construct trainability landscapes over the optimizer learning rate, which affects value learning and actor optimization, and AWR temperature, which controls how selectively the actor imitates high-advantage transitions. Across AntMaze, Cube, and Scene, we observe distinct regimes: high-scoring methods may be broadly accessible or brittle, while broad relative basins may still sit below low absolute ceilings. To interpret these differences, we pair landscapes with post-hoc diagnostics of future-vs-random goal discrimination and AWR weight concentration. % whether the learned advantage function separates matched future goals from random goals, and whether AWR spreads weight across many behavior-supported transitions or concentrates on a small tail. Their relationship to downstream success is task-dependent. On AntMaze, where future goals align with path-like progress, these diagnostics explain landscape regimes. On Cube and Scene, goal ranking and manipulation control decouple: methods can rank goals well while failing downstream, or succeed through action-conditioned advantages despite weak future-vs-random separation. These results show that peak tuned success alone does not establish broadly extractable goal-conditioned behavior. Trainability landscapes expose this gap, while extraction diagnostics offer a lower-cost lens on how learned signals become policies.

The standard post-training recipe for large multimodal models (LMMs) applies supervised fine-tuning (SFT) on curated demonstrations followed by reinforcement learning with verifiable rewards (RLVR). However, SFT introduces distributional drift that neither preserves the model's original capabilities nor faithfully matches the supervision distribution. This problem is further amplified in multimodal reasoning, where perception errors and reasoning failures follow distinct drift patterns that compound during subsequent RL. We introduce \textbf{PRISM}, a three-stage pipeline that mitigates this drift by inserting an explicit distribution-alignment stage between SFT and RLVR. Building on the principle of on-policy distillation (OPD), PRISM casts alignment as a black-box, response-level adversarial game between the policy and a Mixture-of-Experts (MoE) discriminator with dedicated perception and reasoning experts, providing disentangled corrective signals that steer the policy toward the supervision distribution without requiring access to teacher logits. While 1.26M public demonstrations suffice for broad SFT initialization, distribution alignment demands higher-fidelity supervision; we therefore curate 113K additional demonstrations from Gemini~3 Flash, featuring dense visual grounding and step-by-step reasoning on the hardest unsolved problems. Experiments on Qwen3-VL show that PRISM consistently improves downstream RLVR performance across multiple RL algorithms (GRPO, DAPO, GSPO) and diverse multimodal benchmarks, improving average accuracy by +4.4 and +6.0 points over the SFT$\rightarrow$RLVR baseline on 4B and 8B, respectively.

Equipping VLM agents with world modeling capabilities has shown strong potential for complex reasoning and long-horizon planning, while reducing the dependence of policy learning on costly real-world interactions. Existing methods mainly rely on prospective simulation to predict the consequences of candidate actions. However, this forward-only paradigm focuses on *what will happen next* and provides limited constraints for verifying whether an action is causally consistent with the observed state transition, which can lead to plausible-looking but physically incoherent behaviors. In this paper, we challenge the view of world modeling as only prospective prediction and introduce Retrospective World Modeling, a new agent learning paradigm that enables agents to reason backward by estimating the retrospective attribution distribution $P(\hat{a}_{t}|s_t, s_{t+1})$ for the action that most likely caused a given transition. Based on this capability, we formulate the Self-Consistency Reward (SCR), an intrinsic signal that measures the probabilistic consistency between the policy action and the retrospective explanation. Integrating SCR into reinforcement learning provides dense transition-level feedback and steers agents toward behaviors that are both task-effective and physically grounded. Extensive experiments across diverse agentic tasks show that our method substantially improves policy robustness and generalization over prospective-only world modeling baselines.

Deep search agents operate over trajectories spanning dozens of steps, yet standard reinforcement learning provides only a single outcome reward per trajectory — supervision that is far too sparse for effective credit assignment. On-policy self-distillation (OPSD) addresses this by using the model's own logits as dense token-level teachers, but extending it to search agents introduces a fundamental tension: the teacher, having access to privileged information such as the correct answer, produces a distribution that differs systematically from the student's exploration-based reasoning, and naive distillation causes the student to inherit this information asymmetry rather than learn better search strategies. We resolve this tension through two contributions. First, we construct Evidence Anchors — concise, step-level evidence snippets extracted from the web — as privileged information that captures key reasoning steps without revealing the entire answer path. Second, we propose Step-Level Self-Distilled Policy Optimization (SSPO), which converts teacher–student disagreement into step-level advantage weights within GRPO, applied exclusively to incorrect trajectories. This design decouples what to update from how much to update: the outcome reward determines the direction of policy change, while the teacher modulates its magnitude at each step. Correct trajectories are left untouched, preserving their diversity. On Qwen3-8B, SSPO consistently outperforms GRPO across BrowseComp, GAIA, and FRAMES, surpassing or matching GRPO trained with twice as many gradient steps while adding only about 5\% overhead per step from a single additional forward pass.

Large language model (LLM) web agents are usually deployed as tool callers: each turn, the model reads a fresh page observation and emits one structured tool action. When every action is a low-level primitive, horizons grow quickly and so do policy-facing LLM completions, dominating latency and cost on benchmarks such as Mind2Web and WebArena. Recent systems therefore wrap repeated interaction fragments as web skills: callable tools built from successful trajectories or induced programs, so one call can replace several primitives. However, prior skill libraries are still triggered mainly by instruction similarity or coarse site metadata, which yields low skill reuse on held-out sites and leaves much of the potential step and token reduction on the table. We present SkillMigrator, an agent that learns reusable web skills and transfers them across sites by matching layout structure rather than specific element references. Each induced skill is stored as a transferable interaction pattern (TIP): the skill paired with a structural sketch of the snapshot at induction time. At test time, SkillMigrator retrieves TIPs by layout similarity and grounds their references on the live page. The rest of the stack is standard: accessibility-snapshot observations with stable references, and fixed tool calling over primitives plus skill invocations. Compared with the state-of-the-art approaches, SkillMigrator reduces the average LLM-action count on successful trajectories by 8–10\% across both WebArena and Mind2Web at matched success rate.

Reinforcement learning (RL) is now standard for post-training large language models. The same reward optimization that elicits useful capabilities, however, can also reinforce latent backdoors planted earlier in the pipeline: attack success rates under 0.2% after SFT rise to 40--98% on training-distribution inputs after RL, and generalize to 13--30\% on held-out evaluation, with no modification to the RL data, reward function, or training loop. We study this phenomenon in the context of agent models, where SFT poisoning teaches the victim model (Qwen3-8B and a small frontier model) to call an attacker-controlled oracle. Because the oracle can be set up to always return correct answers, calling it earns higher reward than the model's own attempts, and RL reinforces the behavior. We further show that the oracle's RL-time responses can instill persistent biases, such as brand preferences, that survive into deployment even when no tool calls happen. Such patterns appear difficult to detect with current tools: traditional guard models do not flag this mechanism, and LLM-based auditors remain unreliable, achieving only 4.5\% precision even after iterative prompt refinement. Our findings point to the value of post-RL safety evaluation, in particular scrutinizing tool-call patterns such as unnecessary external invocations.