NeurIPS 2026 论文列表

Reinforcement learning (RL) can be used to improve the policy (denoiser) of diffusion large language models (dLLMs), while being hindered by the intractability of the policy likelihood. A dominant and efficient family of methods replaces the likelihood in standard RL with its evidence lower bound (ELBO), estimated from randomly masked sequences. Despite being well-aligned with pre-training, these approaches introduce bias from training-inference mismatch by applying the likelihood surrogate ELBO, which can potentially cause training collapse. In this work, we propose **Guided Denoiser Self-Distillation (GDSD)** to directly distill the denoiser of dLLMs from an advantage-guided self-teacher, derived from the closed-form optimum of reverse-KL regularized RL. GDSD matches the dLLM's denoiser logits to the teacher's via a normalization-free objective, which reduces RL to likelihood-free self-distillation and thus bypasses the TIM biases. Recent ELBO-based methods emerge as instances of applying different distillation divergences, but with diagnosable pathologies that GDSD avoids. On planning, math, and coding benchmarks with LLaDA-8B and Dream-7B, GDSD consistently outperforms prior state-of-the-art ELBO-based methods with a more stable training reward dynamics, achieving test-accuracy improvements of up to $+19.6\\%$. These results suggest that direct denoiser self-distillation, without relying on ELBO likelihood surrogate, can provide a more stable and effective RL procedure for dLLMs.

Diffusion distillation, exemplified by Distribution Matching Distillation (DMD), has shown great promise in few-step generation but often sacrifices quality for sampling speed. While integrating Reinforcement Learning (RL) into distillation offers potential, a naive fusion of these two objectives relies on suboptimal raw sample evaluation. This sample-based scoring creates inherent conflicts with the distillation trajectory and produces unreliable rewards due to the noisy nature of early-stage generation. To overcome these limitations, we propose GDMD, a novel framework that redefines the reward mechanism by prioritizing distillation gradients over raw pixel outputs as the primary signal for optimization. By reinterpreting the DMD gradients as implicit target tensors, our framework enables existing reward models to directly evaluate the quality of distillation updates. This gradient-level guidance functions as an adaptive weighting that synchronizes the RL policy with the distillation objective, effectively neutralizing optimization divergence. Empirical results show that GDMD sets a new SOTA for few-step generation. Specifically, our 4-NFE (Number of Function Evaluations) models outperform the quality of their multi-step teacher and substantially exceed previous DMDR results in GenEval and human-preference metrics, exhibiting strong scalability potential.

Off-policy reinforcement learning (RL) for large language models bottlenecks at the rollout stage: every parameter update invalidates the key-value (KV) cache of in-flight agent trajectories, forcing either expensive recomputation or staleness that destabilizes training. Recent advances such as M2PO and asynchronous RLHF tolerate token-level staleness by reweighting the policy gradient with second-moment importance corrections, but they implicitly assume that a stale rollout is sampled from a well-defined behavior policy $\pi_{\theta_{\text{old}}}$. We show that this assumption is silently violated by every modern partial-rollout system: once a stale KV cache is reused under updated parameters $\theta_{\text{new}}$, the resulting token distribution is a hybrid $\pi_{\text{hyb}}$ whose attention projections consume stale keys and values while its MLP, output, and gating projections use the new weights. The token-level importance weight $w_t = \pi_{\theta_{\text{new}}}(y_t \mid x_{<t}) / \pi_{\theta_{\text{old}}}(y_t \mid x_{<t})$ corrects the sampling distribution but not this hidden-state-level mismatch, leaving an uncorrected source of policy-gradient bias that grows with both staleness and weight-update magnitude. We introduce Gradient-Coupled KV Drift (GCD), an exception-safe correction comprising (i) a high-probability concentration bound on the drift, with explicit per-layer terms for full-attention and Gated DeltaNet layers; (ii) a tail-event detector that flags inputs outside the bound’s regime of validity; (iii) a composite safety guarantee that treats the bias correction as trustworthy only when both the analytical estimate and the detector agree; and (iv) a variance-preservation theorem showing that GCD composes with the M2PO trust region without re-tuning. On Qwen3.5-9B trained with GRPO on MATH-500 and $\tau$-bench, GCD enables stable training at staleness 512 (a regime where M2PO collapses), reduces rollout wall-clock by 38\% versus full recomputation, and lifts $\tau$-bench retail pass$^4$ by 5.7 points over the M2PO baseline.

Fine-tuning diffusion models via online reinforcement learning (RL) has shown great potential for enhancing text-to-image alignment. However, since precisely specifying a ground-truth objective for visual tasks remains challenging, the models are often optimized using a proxy reward that only partially captures the true goal. This mismatch often leads to reward hacking, where proxy scores increase while real image quality deteriorates and generation diversity collapses. While common solutions add regularization against the reference policy to prevent reward hacking, they compromise sample efficiency and impede the exploration of novel, high-reward regions, as the reference policy is usually sub-optimal. To address the competing demands of sample efficiency, effective exploration, and mitigation of reward hacking, we propose \textbf{G}ated and \textbf{A}daptive \textbf{R}egularization with \textbf{D}iversity-aware \textbf{O}ptimization (\textbf{GARDO}), a versatile framework compatible with various RL algorithms. Our key insight is that regularization need not be applied universally; instead, it is highly effective to selectively penalize a subset of samples that exhibit high uncertainty. To address the exploration challenge, GARDO introduces an adaptive regularization mechanism wherein the reference model is periodically updated to match the capabilities of the online policy, ensuring a relevant regularization target. To address the mode collapse issue in RL, GARDO amplifies the rewards for high-quality samples that also exhibit high diversity, encouraging mode coverage without destabilizing the optimization process. Extensive experiments across diverse proxy rewards and hold-out unseen metrics consistently show that GARDO mitigates reward hacking and enhances generation diversity without sacrificing sample efficiency or exploration, highlighting its effectiveness and robustness.

Both full fine-tuning (Full FT) and parameter-efficient methods like LoRA add weight updates without regard to the spectral structure that pretraining has established. This allows noisy gradients from a small fine-tuning distribution to freely perturb the robust features learned through pretraining. We first identify *spectral preconditioning* as the key missing ingredient: reparameterizing each weight $\mathbf{W}$ through its full-rank SVD and freezing one singular basis confines every update to the pretrained column space, yielding a preconditioned optimizer that outperforms unconstrained Full FT at the same parameter count. To make this insight practical, we propose FuRA (**Fu**ll-**R**ank **A**daptation), which factorizes $\mathbf{W}$ via a block tensor-train decomposition $\mathbf{W}=\mathbf{L}\mathbf{S}\mathbf{R}$: the large core $\mathbf{L}$ is frozen at the pretrained block-wise SVD basis while only the small core $\mathbf{R}$ and per-block singular values $\mathbf{S}$ are trained. This single design choice simultaneously delivers full-rank spectral preconditioning, full-rank update capacity, and parameter, step time, memory efficiency on par with LoRA. FuRA outperforms Full FT on LLM fine-tuning ($+1.37$ on LLaMA-3-8B commonsense reasoning), LLM math reinforcement learning, and VLM visual instruction tuning. The 4-bit quantized version QFuRA also outperforms QLoRA.

Vision-language models (VLMs) have achieved remarkable progress in 2D visual understanding tasks, yet the passive view-centric perception paradigm limits their capacity for 3D spatial reasoning. Such reasoning often requires actively acquiring spatial evidence beyond the currently visible views, such as camera poses and cross-view spatial relationships, to infer 3D relationships that are not explicitly represented in 2D images. However, existing methods either lack active exploration or merely provide additional views during inference, without explicitly acquiring such spatial evidence beyond 2D images. In this paper, we propose View-to-World (V2W), an active exploration framework that enables VLMs to actively explore 3D worlds and lifts VLMs from passive view-centric perception to active world-centric spatial reasoning. Given a spatial reasoning task with multi-view images, V2W first reconstructs an explicit 3D world with geometry, camera poses, and language-grounded semantics, and exposes it to VLMs through visual and linguistic interfaces. Through multi-turn interaction with these interfaces, VLMs iteratively acquire spatial evidence from the constructed 3D world, enabling active world-centric spatial reasoning. V2W can be integrated with existing VLMs in a training-free manner and can be further enhanced by optimizing the exploration policy via reinforcement learning. Experiments on spatial mental modeling benchmarks demonstrate that V2W substantially improves VLM baselines in the training-free setting and achieves state-of-the-art performance with the learned exploration policy, highlighting the importance of active exploration for spatial intelligence. The code will be available upon acceptance.

Text-to-image prompt optimization is often treated as prompt rewriting: given a user prompt, a language model expands it with richer style, composition, and photographic details. While this paradigm can improve visual appeal, it may silently weaken the user's actual intent, especially when the prompt specifies fine-grained objects, attributes, and relations. We argue that faithful prompt optimization requires a different objective: improving perceptual quality while preserving the compositional structure that makes the prompt semantically correct. We propose \ours{}, a single-step reinforcement learning framework that transforms noisy scene-graph feedback into a learnable prompt-editing policy. Instead of freely rewriting the entire prompt, \ours{} first extracts a calibrated scene graph from the input and then performs conservative deletion, reordering, and insertion actions conditioned on both text and graph representations. This design exposes structured intent to the optimizer while restricting unnecessary prompt drift. To make object--relation feedback usable for policy learning, \ours{} replaces sparse thresholded grounding rewards with uncertainty-aware smooth rewards and estimates advantages with environment-aware grouping over multiple image samples from the stochastic text-to-image generator. Across prompt-optimization datasets and generation backbones, \ours{} improves the faithfulness--aesthetics trade-off, with particularly consistent gains in relational fidelity where generic rewriting and aesthetics-oriented optimization often fail. We further evaluate with both human preference and Gemini-VQA, showing that the improvements are not confined to the SG reward used during training. Appendix additionally compares against a high-budget iterative GPT-4o structural-feedback baseline to contextualize the efficiency of amortized policy learning. More broadly, our results suggest that structural constraints should not remain only post-hoc evaluators of generated images; when properly calibrated and stabilized, they can be internalized as training signals for reliable, intent-preserving prompt policies.

Scaling robot demonstrations does not guarantee better imitation: redundant trajectories dilute learning, while rare, imperfect segments may encode critical transitions. Prior curation methods overlook such dependencies by scoring samples independently at the trajectory or segment level. We reframe robot demonstration curation as a coverage-aware sequential selection problem and propose PROSE, which selects data by maximizing marginal utility under a fixed budget. PROSE unifies three signals: (i) influence on closed-loop return, (ii) kinematic reliability, and (iii) coverage gain measuring novelty with respect to the retained set, allowing full trajectories and fine-grained segments jointly participate in scoring and filtering. Across three Robomimic simulation tasks and three real Franka tasks, PROSE achieves the best success rate on every task we evaluate against three state-of-the-art curators (CUPID, SCIZOR, Demo-SCORE), with average closed-loop gains of +15\% in simulation and +35\% in real-world. Ablations attribute the gain primarily to the coverage step.

Reinforcement learning from verifiable rewards (RLVR) has shown strong promise for LLM reasoning, but typical outcome-based RLVR methods remain inefficient on hard problems. Correct final-answer rollouts are rare, and standard sample-level credit assignment fails to leverage partial reasoning progress embedded in unsuccessful attempts. To address this problem, we introduce **SCRL** (**S**ubproblem **C**urriculum **R**einforcement **L**earning), a curriculum reinforcement learning framework built on verifiable subproblems derived from reasoning chains. Given a reference solution, SCRL derives a series of verifiable subproblems and constructs a subproblem curriculum, with the final subproblem fixed as the original problem. This converts partial progress on hard problems into verifiable learning signals. Algorithmically, we propose *subproblem-level normalization*, a training technique based on RLVR that normalizes rewards independently at each subproblem position within the rollout group. By assigning the resulting advantages to the corresponding answer spans, we enable finer-grained credit assignment without external rubrics or reward models. Our theoretical analysis shows that this subproblem curriculum makes hard problems more learnable by lifting them out of gradient dead zones, with larger relative gains as the original problem becomes harder. Across seven mathematical reasoning benchmarks, SCRL outperforms strong curriculum-learning baselines, yielding +4.1 and +1.9 average-point gains compared to GRPO on Qwen3-4B-Base and Qwen3-14B-Base respectively. On three hard benchmarks (AIME24, AIME25, and IMO-Bench), SCRL further yields point gains of +3.7 in pass@$1$ and +4.6 in pass@$64$ on Qwen3-4B-Base, suggesting improved exploration on hard reasoning problems. Further ablations show that the proposed credit assignment is effective and that the gains do not require highly curated subproblems or strong external generators, highlighting SCRL as a practical curriculum framework that enables fine-grained credit assignment for LLM reasoning.

Model-based reinforcement learning in partially observable environments requires inferring hidden states, learning latent dynamics, and planning under the resulting belief. Practical algorithms often solve one aspect of the above pipeline while relying on approximations for the others that are often poorly understood, while existing POMDP theory typically analyzes computationally inefficient theoretical algorithms far from modern deep RL practice. To fill this gap, we propose a practical end-to-end algorithm that approximately solves the hidden state filtering problem via sequential Monte Carlo (SMC), updates a latent dynamics model online, and performs approximate online planning for $d$ steps before bootstrapping leaves with a state-value critic (a QMDP approximation). We analyze each approximation and prove an end-to-end regret bound decomposing the error into model estimation error, particle filter error, Monte Carlo rollout error, and a structural bias term induced by the QMDP approximation controllable by the planning depth that we further analyze in various special cases. The resulting guarantee makes explicit how statistical sample size, particle budget, rollout budget, and planning depth affect performance. We instantiate our algorithm in an MPC-based deep RL framework, demonstrating its practicality on a variety of environments. Our results provide a practical and theoretically grounded route for bridging theory and practice within online decision-making in POMDPs.

Large Language Model-based Automated Algorithm Design (LLM-AAD) has shown promising results across diverse domains by coupling LLMs with iterative search. Some recent efforts have begun to fine-tune LLMs for algorithm design through reinforcement learning. However, these attempts still rely on outcome-level supervision, leaving intermediate ideas and plans without direct credit, even though they determine the conceptual direction of algorithm design. In this paper, we propose Algorithm Tree Policy Optimization (ATPO), which utilizes tree-structured rollouts to decouple the generation of conceptual ideas from their specific code implementations. By applying group-relative advantage estimation, ATPO assigns credit to both strategic ideas and algorithmic implementations, thereby optimizing the policy across multiple levels of abstraction. Integration with FunSearch, EoH, and OpenEvolve demonstrates improvements over their original counterparts. Notably, the learned policy transfers effectively to unseen problem instances and alternative search methods, indicating that it can improve the generalization of algorithm design policies.

Existing approaches to optimization modeling using large language models (LLMs) treat tasks through from-scratch construction, generating variables, objectives, and constraints without leveraging reusable structural knowledge. However, in real-world scenarios, expert modelers typically begin by identifying the canonical type or core structure of a problem, followed by iterative refinements based on specific task requirements. To incorporate this expert incremental logic, we propose a novel two-stage framework. In the first stage, we conduct Logic-Anchored Data Synthesis starting from 17 canonical prototypes, preserving critical bottleneck constraints and generating synthetic problem–formulation pairs. These pairs, together with the OR-Instruct dataset, are used for supervised fine-tuning (SFT) to initialize a policy. Separately, for each generated problem, we produce multiple modeling trajectories that follow a prototype-grounded order with variables preceding expressions. These trajectories are ranked according to criteria of accuracy and efficiency, and subsequently used to train a Logical Reward Model (LRM). In the second stage, we introduce Logic-Test-Time Group Relative Policy Optimization (Logic-TGRPO), which leverages the LRM during test-time reinforcement learning to reward accurate prototype identification and disciplined structural adherence while penalizing illogical patterns. Evaluated across seven benchmarks, our 8B parameter model achieves an average accuracy of 79.9%, outperforming comparably scaled baselines and competing effectively with heavyweight multi-agent systems with far fewer parameters. Strong performance on out-of-distribution tasks confirms that the model exhibits genuine incremental reasoning rather than mere memorization. Ablation studies further validate that both stages in our framework are essential.

Deep narrow neural networks underlie modern architectures from ResNets to Neural ODEs, yet their fundamental computational properties at constant width remain poorly understood. We prove that, for every deterministic Turing machine $M$, there is a single $ReLU$ network $N_M \colon [0,1]^4 \to [0,1]^4$ of width $4$ and depth $\mathrm{poly}(|M|)$, polynomial-time constructible, whose iterates simulate $M$ on every infinite binary tape encoded in a \emph{convex} initial region of $[0,1]^4$, with iteration count proportional to Turing-machine time. This establishes a tight correspondence between depth in width-$4$ networks and time on a Turing machine. The correspondence has three immediate consequences. First, verification of width-$4$ $ReLU$ networks spans the full complexity hierarchy along the single axis of depth: it is $NP$-complete for unary-encoded time bound, $NEXP$-complete for binary-encoded time bound, and $RE$-complete (undecidable) for unbounded time. Second, the hardness extends with two extra neurons of width to every common smooth activation: at width $6$, the robust verification problem (a promise problem with $\delta/2\delta$ gap) is $NP$-complete for $\tanh$, sigmoid, GELU, Swish, ELU, and any polynomial-time-computable continuous non-polynomial activation admitting a polynomial-time computable Lipschitz constant. A corollary is a discretized Neural ODE analogue at dimension 5. The simulation and statements are carried out on convex inputs, covering a regime previously inaccessible because of a continuity obstruction inherent to $\ReLU$ networks.

Continual Reinforcement Learning (CRL) requires agents to adapt to sequentially arriving tasks while retaining performance on previous ones. Most existing CRL methods rely on model-free frameworks, addressing catastrophic forgetting via regularization, parameter isolation, or experience replay. World models offer a promising alternative by accumulating transferable dynamics knowledge and enabling planning-based decision-making. However, existing world model approaches suffer from slow adaptation at task boundaries: abrupt reward shifts mislead the planner, fixed planning horizons accumulate prediction errors when the dynamics model is unreliable, and quality-agnostic replay dilutes training while exacerbating forgetting. We propose HERD (Horizon-adaptive Elite Replay with reward Disagreement), which addresses these challenges from two complementary perspectives: for $\textbf{\emph{how to learn}}$, Uncertainty-aware Reward Planning leverages ensemble disagreement to accelerate reward calibration and guide exploration, while Adaptive Planning Horizon adjusts rollout depth based on dynamics prediction error to prevent error accumulation; for $\textbf{\emph{what to learn}}$, Elite Experience Replay applies quality-stratified reweighting of historical trajectories to strengthen knowledge retention. Together, these components $\textbf{\emph{accelerate new-task adaptation and mitigate catastrophic forgetting.}}$ Experiments on Continual World, Continual Bench, and DM Control show that HERD consistently outperforms existing methods in overall performance, forgetting mitigation, and new-task adaptation.

In real-world tasks such as robotic control and autonomous driving, agents often face non-Markovian environments due to incomplete observations. Solving such information-impoverished decision tasks requires reconstructing the context from history. However, existing recurrent encoders have two fundamental limitations: First, the model suffers from a Markovian compression bottleneck, i.e., it only processes observation history as input but neglects thought process history. This leads to the loss of deep semantic cues from early interactions, which are easily overwritten by noise over time due to recursion. Second, it lacks the adaptive capability to adjust the memory span for different tasks. To address these limitations, we propose FracEncoder, a context encoder based on fractional-order neural differential equations. First, FracEncoder leverages the nonlocal property of the fractional-order derivative and uses a power-law kernel to directly incorporate the trajectory of the thought process into the evolution of the current state, effectively overcoming the temporal locality limitations of traditional models. Second, the fractional order $\alpha$ is set as an end-to-end learnable parameter. It explicitly characterizes how strongly the system anchors to its history, so the model can discover a suitable memory span for each task. We evaluate FracEncoder on more than ten benchmarks across three settings, namely partially observable tasks, meta reinforcement learning, and delayed-observation control. FracEncoder consistently matches or outperforms representative baselines, including GPIDE, GRU-ODE, ReSeL, and LRU. On the challenging 8-step delayed-observation task, it reaches average returns of $-511.5 \pm 7.9$ on Pendulum and $732.1 \pm 152.7$ on Hopper. As a plug-in module, FracEncoder also yields stable gains across different reinforcement learning algorithms. Ablation studies further show that the learned $\alpha$ aligns with the intrinsic memory demand of each task, providing an interpretable physical measure of the cognitive complexity of non-Markovian tasks. The code is available at https://anonymous.4open.science/r/FracEncoder-60BA.

Accurate 3D surface reconstruction from unposed images remains a fundamental challenge in computer vision. Recent feed-forward pointmap methods eliminate the need for camera poses but inherently produce discrete, sparse representations. In this paper, we present a generalizable pipeline that directly recovers continuous, high-fidelity 3D surfaces from unposed images by seamlessly integrating pointmap- and ray-based representations. Our approach first extracts explicit geometric priors using a pre-trained pointmap model. A query-view-centric selection module then identifies geometrically relevant reference views via a robust surfel voting mechanism. Next, a cross-attention regressor bridges target query rays with reference features to estimate a continuous coarse surface. Finally, a raylet-based module aggregates local 3D features to carve out high-frequency details and analytical normals. Trained as a single foundation model on 9 diverse datasets, our method demonstrates exceptional zero-shot generalization on 6 unseen datasets, clearly surpassing state-of-the-art baselines in geometric accuracy and detail preservation.

Active feature acquisition (AFA) requires an agent to choose a budget-limited set of costly features before making a prediction. Greedy conditional-mutual-information (CMI) and value-of-information policies perform well when useful features are individually informative, but fail under joint-only evidence: a query is valuable only because it enables a future evidence set. To address this, we propose Forward-$\Phi$, a model-based non-myopic acquisition scorer that requires a known structural causal model (SCM) or generative AFA model. Forward-$\Phi$ evaluates a candidate query by its Shapley contribution to budget-feasible future action sets in a cooperative game on the expected forward utility. Two variants share the same game: a raw score that keeps both singleton and interaction value, and a residual score that strips the singleton component to isolate joint-only utility on diagnostic probes. On AFABench CUBE-NM, the raw Monte Carlo (MC) Forward Shapley scorer reaches $0.664 \pm 0.024$ accuracy over five MC seeds, well above all reported baselines including the OL-MFRL ($0.235$) and ODIN-MFRL ($0.160$) reinforcement-learning loops trained under the same split and shared pretraining protocol. Controlled SCM and Bayesian-network (BN) probes confirm the predicted mechanism: Forward-$\Phi$ yields substantial gains when target information is jointly carried, and matches Greedy on greedy-aligned controls. The primary contribution is the belief-conditional Forward Shapley state-action score, a credit-assignment signal derived from Shapley axioms for non-myopic AFA under joint-only evidence.

The uncertainty of a statistical model is most commonly factorized into an aleatoric and an epistemic part. This factorization changes how predictions are interpreted in downstream decision tasks. Importantly, except for the idealistic scenario with no model mismatch, the quantification is a characteristic of the model and not the data generating process. In this paper, we propose Forgetting to Improve a method that reduces this discrepancy by incorporating the task into the modeling framework. Our key insight is to acknowledge that in scenarios of model mismatch, data can have a detrimental effect on the modeling for a specific task. Based on this insight, we propose an influence function for Gaussian process models that allows for principled removal of detrimental data samples. We showcase the flexibility of this approach by demonstrating significant improvements across a range of tasks, including Bayesian optimization, model-based reinforcement learning, and transductive learning.

Cooperative multi-agent reinforcement learning (MARL) under sparse rewards remains fundamentally challenging because agents often fail to concentrate their influence, leading to insufficiently coordinated exploration. To address this, we propose the Focusing Influence Mechanism (FIM), a framework that encourages agents to focus their influence on under-explored parts of the state space through an entropy-based criterion, while leveraging eligibility traces to enable multiple agents to consistently align and sustain their influence on the same parts of the state space when beneficial, thereby promoting coordinated and persistent joint behavior. By emphasizing under-explored regions of the state space, FIM facilitates more efficient and structured exploration even under extremely sparse rewards. Across diverse MARL benchmarks, FIM consistently improves cooperative performance over strong baselines.