NeurIPS 2026 论文列表

We present Reflow-regularized Flow Matching Policy Gradients (ReFPO), a simple online RL method that adds explicit Reflow regularization to FPO for efficient flow-based control. We uncover a key structural property: the gradient updates in Flow Matching Policy Gradients (FPO) can be interpreted as an implicit advantage-weighted Reflow process, providing a new geometric perspective on flow-based policy gradients. Building on this insight, ReFPO introduces an explicit geometric regularizer that can be implemented with a single line of code change without incurring additional computational overhead or auxiliary distillation stages. By synergizing advantage-guided updates with path rectification, our method reduces CFM proxy-ratio spikes, stabilizes PPO-style training, and enables high-fidelity one-step inference that often matches or exceeds multi-step performance. We experimentally demonstrate that ReFPO improves average performance and discretization robustness across GridWorld, MuJoCo Playground, and high-dimensional Humanoid Control tasks, providing a scalable and stable approach for generative policies in complex physical simulations.

In retrieval-augmented generation (RAG), knowledge conflicts arise when retrieved contexts disagree with each other or with the model's parametric memory. Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a standard approach to address such conflicts. However, existing RLVR-based methods are typically formulated under a single-conflict assumption. As a result, the type with stronger advantage signals dominates policy updates, leaving the other under-optimized and thus yielding imbalanced performance across conflict types. To address this challenge, we propose ReCon, a type-aware RLVR framework for balanced learning across Inter-Context and Context-Memory conflicts. Specifically, ReCon computes a step-level imbalance signal from type-wise cumulative advantages and triggers resampling on the weaker conflict type. Instead of resampling from scratch, ReCon branches from informative failures. It selects an informative failed trajectory using uncertainty and closeness signals, then branches at an entropy-jump point, concentrating additional rollouts on critical reasoning decisions. Experiments show that ReCon yields stronger and more balanced performance across knowledge-conflict and multi-hop QA benchmarks, and also generalizing well under mixed-conflict settings.

Pre-training large language models on massive GPU clusters has made hardware faults routine rather than rare, driving the need for resilient training systems. Yet existing frameworks either focus on specific parallelism schemes or risk drifting away from a failure-free training trajectory. We propose ReCoVer, a resilient LLM pre-training system that upholds a single invariant: each iteration keeps the number of microbatches constant, ensuring per-iteration gradients remain stochastically equivalent to a failure-free run. The framework is organized as three decoupled protocol layers: (1) Fault-tolerant collectives that isolate faults from propagating across replicas; (2) in-step fine-grained recovery that preserves intra-iteration progress and prevents gradient corruption; (3) versatile-workload policy that dynamically redistributes microbatch quotas across the survivors. The design is parallelism-agnostic, integrating directly with both 3D parallelism and Hybrid Sharded Data Parallel (HSDP) as a drop-in substrate. We evaluate our implementation on end-to-end pre-training tasks for up to 512 GPUs, ReCoVer successfully preserves the training trajectory from a failure-free reference despite of 256 GPUs lost spread across the run. For comparison with checkpoint-and-restart baselines, ReCoVer demonstrates $2.23\times$ higher effective throughput after successive failures. This advantage results in ReCoVer processing 74.9% more tokens at 234 GPU-hours, with the gap widening as the training prolongs.

Comics and manga tell stories through characters whose appearances change from panel to panel, yet readers recognise them hundreds of pages later without ever consulting a cast list. Character re-identification models rarely face the same challenge: existing evaluations give them a fixed set of identities before the story begins. **Re:Cognize** instead asks models to recognise characters as the story unfolds and learn who is who along the way. Across three manga corpora, we evaluate five Re-ID backbones and **MECHA**, a memory-augmented baseline we introduce, and find that the central difficulty lies somewhere unexpected. Recognising familiar characters is nearly solved. Knowing which matches to trust is not. When models build a gallery and grow it from their own predictions, their performance deteriorates, even though correct identity labels would improve it by over twenty points. This raises a simple question: when should a model trust its own match? We derive a single comparison that determines when gallery updates improve recognition and use it to design **VOUCH**, which updates its memory only when the page provides supporting evidence. Without fitting additional parameters, VOUCH recovers between one-third and two-thirds of the improvement available from perfect labels. RE:COGNIZE measures whether a model can recognise characters as it reads; VOUCH shows how to remember them reliably. We focus our methodological claims on identity maintenance for previously encountered characters, while separately evaluating new-character emergence as a diagnostic under a fixed reference rule to distinguish recognition from discovery.

Masked diffusion language models (dLLMs) have recently emerged as a competitive alternative to autoregressive language models, with the promise of faster inference via parallel token generation. A notable limitation of the masked formulation, however, is that once a token has been unmasked it can no longer be revised, leaving dLLMs vulnerable to early sampling mistakes. To address this, a growing body of work has sought to extend masked dLLMs with self-correcting (remasking) capabilities. One appealing subset of these methods does so in a training-free, post-hoc manner based on token confidence, with encouraging early reported results. In this work, we revisit the empirical evaluation of a representative post-hoc remasking method, WINO, and find that under standard decoding settings (shorter block lengths) it brings little-to-no benefit over confidence-based unmasking alone. Extending the evaluation to non-greedy decoding, we find that while confidence-based remasking can mitigate errors introduced by increased stochasticity to some extent, it also exacerbates the diversity collapse previously reported for confidence-based unmasking. Overall, our results show that the benefits of post-hoc confidence-based remasking are highly setting-dependent, underscoring the need for a more comprehensive evaluation framework.

Recent 3D foundation models, such as DUSt3R, MASt3R, VGGT, $\pi^3$, and Depth Anything 3, provide strong feed-forward depth and pose estimates on pinhole imagery, but degrade sharply under fisheye camera geometry. We show that this failure is partly caused by a pinhole camera bias in the positional encodings of pretrained 3D foundation models, and propose RayTun3R, a lightweight camera adaptation approach. It keeps the pretrained network fixed and adapts only lightweight components tied to token position and camera geometry. RayTun3R learns parameter-efficient residual corrections to absolute and rotary positional encodings, together with parameter-free tokenization and corrections to prediction-grid coordinates that remove residual pinhole assumptions. The resulting adapter contains only 10,752 trainable parameters and can be learned from a short temporal segment using geometric losses. Once adapted, RayTun3R transfers effectively to the remaining frames of the sequence without incurring additional runtime costs. Across diverse fisheye datasets with fields of view from $110^\circ$ to $200^\circ$, our adapter reduces rotation error by $2-12\times$ relative to the unadapted model, outperforms LoRA while using $\sim14\times$ fewer trainable parameters, improves pose over adaptation-free baselines while avoiding their multi-view inference cost, and remains competitive on depth accuracy.

Time series foundation models (TSFMs) commonly adapt to new data by attaching a single trainable head to a frozen backbone, a one-size-fits-all setup that underfits heterogeneous regimes. Replacing the head with a mixture of experts is the standard upgrade, but on instance-normalized backbones (the dominant TSFM design class) it fails: routing entropy collapses to zero and one expert absorbs every input, a failure we call *normalization-induced routing collapse*. Standard MoE rescue mechanisms do not repair it, because the cause is in the router's input, not its optimization. Pre-encoder normalization strips the statistics a router would need to tell regimes apart. A mutual-information decomposition makes this precise and yields a signal-ratio that, computed before training, predicts dataset vulnerability (Spearman $\rho = -0.88$). Eight causal controls, including a vision-modality replication, isolate instance normalization as the cause. The prescription is a minimal causal intervention: *Raw-Routed Mixture of Adapters* (RR-MoA), which routes on the raw, pre-normalization input. Under a strictly frozen backbone, RR-MoA wins 54/54 comparisons against the strongest fixed adapter and significantly outperforms LoRA, TRACE, AdaMix, and full fine-tuning. The effect generalizes across six backbones and an imputation task. Frozen RR-MoA also beats full fine-tuning by 12–79% (the *Frozen Paradox*); two architecturally distinct variants confirm the principle generalizes beyond this specific router. Code is provided in the supplementary material.

Long-context recall in linear-time sequence models highlights a tradeoff in how they write to memory. State-based linear models, such as state-space models (SSMs) and linear Transformers, write densely, updating the entire state for each newly arrived token, which leads to interference and makes specific past tokens hard to recover. Sliding-window attention (SWA) exhibits the opposite behavior: it writes sparsely by storing explicit token representations, but only within a fixed window, so recall drops once the relevant token is evicted. Interpolating between these models, we introduce Raven, a linear-time sequence model that maintains a fixed set of memory slots and, at each step, decays and updates only a selected subset via learned, input-dependent routing. This lets Raven mitigate SWA's position-based overwriting and hard eviction while reducing interference from dense state updates in SSMs, thereby preserving long-range content much more effectively. Across recall-intensive benchmarks, Raven is competitive with or outperforms prior linear-time baselines, achieving strong long-context recall where both SWA and SSMs sharply degrade. It remains effective when extrapolating to context lengths as large as 16x its training length, with similar gains in hybrid architectures.

On-policy reinforcement learning has become a central paradigm for improving the reasoning abilities of large language models. However, its effectiveness is often limited by reward sparsity: when a model fails to discover correct trajectories for difficult problems, the optimization process receives little useful signal and may stagnate. Existing approaches mitigate this issue by incorporating off-policy demonstrations, expert traces, or model-generated solutions, but they typically require the auxiliary data to match the format of the reinforcement-learning task, often relying on rejection sampling from stronger models to obtain suitable training trajectories. We introduce Rationale-Guided Policy Optimization (RGPO), a framework that adaptively leverages ground-truth rationale information according to the model’s current capability while preserving its freedom to explore. Rather than treating reference solutions as fixed imitation targets, RGPO uses them as temporary scaffolds: rationales help the model generate improved responses, after which only higher-reward, model-generated solutions are transferred back to the original unguided setting. This design allows training to exploit available ground-truth information without requiring off-policy data to follow the same format as the RL task. Across both language-only and vision-language reasoning settings, RGPO consistently improves performance over RLVR baselines, and ablation studies show that adaptive rationale guidance is a key contributor to these gains. These results suggest that RGPO offers a practical and general approach for reducing reward sparsity, stabilizing reinforcement learning, and improving reasoning performance in both text-only and multimodal models.

Real-world television and streaming pipelines often deliver compressed HD video to increasingly capable displays, making video super-resolution (VSR) both practically important and fundamentally ambiguous. Existing generative VSR methods are largely receiver-only, requiring the model to infer missing high-frequency structures from degraded pixels alone. We propose MetaVSR, a sender--receiver collaborative framework that transmits compact structural edge metadata as bitrate-constrained side information for generative VSR. MetaVSR builds on a one-step video diffusion transformer and encodes both the low-quality video and transmitted metadata through the native 3D VAE, enabling token-level fusion within the original DiT attention blocks without additional control branches. We further introduce a rate-aware Canny edge metadata selection strategy that allocates metadata bits to structurally challenging regions, and evaluate reconstruction quality under matched total transmission budgets that include both compressed video and metadata. Across five public VSR benchmarks, MetaVSR shows that budgeted sender-side structural metadata can improve reconstruction over the receiver-only DOVE-5B baseline, increasing PSNR by 1.03 dB and SSIM by 0.055 on average while using a smaller 2B backbone. On MetaSPS, our rate--distortion evaluation shows approximately 20\% rate saving under light-noise degradation and more than 50\% under heavy-noise degradation compared with a no-metadata generative baseline. These results demonstrate that compact sender-side edge metadata can reduce reconstruction ambiguity and improve bitrate-efficient VSR when its transmission cost is explicitly accounted for.

We revisit the classical problem of learning general Gaussian halfspaces in the presence of label noise, with particular focus on the large-bias regime where information-computation gaps are conjectured in several standard label-noise models. Concretely, we present and analyse the rare-tail mean estimator, an extremely simple one-shot statistic that averages only the observed minority class. In the large bias regime, classical Averaging (Servedio, 1999), and related global first-moment statistics, are provably suboptimal, while the rare-tail mean estimator recovers the target halfspace direction at the information-theoretic sample rate (modulo log factors) under a broad class of label noise models, including probit noise and margin-monotone noise with very slow decay away from the boundary. Notably, our guarantees allow labels arbitrarily close to the decision boundary to be nearly random, yet still obtain strong directional recovery from a small number of samples. As an application, we show that in a two-stage curriculum learning setup, where an average-case margin-dependent noise mechanism precedes the uniform-noise target (the worst-case noise profile in the same family) with the same Bayes optimal separator, our estimator supplies the missing directional accuracy needed by existing refinement routines, yielding an end-to-end sample-optimal learner in the large-bias regime. The estimator also cleanly interpolates between the noiseless and noisy regimes. In low-noise settings, it retains the same sample complexity rates as classical separability-based methods for realizable halfspace learning, while continuing to admit meaningful guarantees even after linear separability has been destroyed by noise.

Rare disease diagnosis is a fundamental challenge due to heterogeneous and overlapping clinical phenotypes. While recent LLM-based agentic systems have demonstrated promise by integrating external medical tools and knowledge, they typically rely on large-scale or commercial LLMs, limiting their practical deployment in resource-constrained settings. In this work, we observe that their performance significantly degrades when using small-scale LLMs, due to entangled workflows and unreliable multi-evidence aggregation. To address this issue, we propose RADAR, a rare disease diagnostic framework designed for small-scale LLMs. RADAR adopts a divide-and-conquer paradigm that decouples heterogeneous diagnostic evidence into specialized workflows, reducing reasoning complexity and improving robustness. It further introduces a knowledge-driven self-evaluation mechanism that performs evidence-aware reasoning over structured disease knowledge to produce interpretable reliability scores by identifying key supporting phenotypes and conflicting evidence. Finally, a multi-evidence fusion strategy integrates outputs from multiple sources based on reliability and clinical agreement, resolving conflicts and producing stable diagnostic rankings. Experiments on three benchmarks show that RADAR consistently outperforms various state-of-the-art baselines, including general and medical LLMs and agentic systems. Code will be released.

The growing context lengths of large language models (LLMs) have made key-value (KV) cache memory a critical bottleneck during inference. Vector quantization (VQ) with commutative codebooks provides an effective Rotary Position Embedding (RoPE)-aware solution by enabling reconstruction-free attention over quantized key indices. However, existing commutative VQ methods still parameterize scalar codebook coefficients across RoPE subspaces and codewords largely independently, leaving parameter-side redundancy underexploited. We propose $\textbf{RankVQ}$, a low-rank parameterized commutative vector quantization method for KV cache compression. Instead of factorizing the deployed codebooks, RankVQ factorizes the underlying scalar coefficient matrices used to construct key and value codebooks. This design preserves the pseudo-symmetric structure required for RoPE commutativity on the key side, while improving the quality and compactness of offline codebook construction. The low-rank factors are discarded after codebook construction, so RankVQ keeps the same online inference form as standard commutative VQ and introduces no additional decoding overhead. Experiments on LongBench, InfiniteBench, and GSM8K show that RankVQ achieves a stronger accuracy--compression trade-off than directly comparable KV quantization baselines, with particularly clear gains in the 1-bit regime.

Discrete autoregressive (AR) text-to-image (T2I) models adopt a two-stage paradigm in which a VQ tokenizer maps images to discrete codes and an AR policy models their distribution. Current post-training methods optimize only the AR policy while keeping the VQ decoder frozen. We show that this practice introduces Latent Covariate Shift: as the policy evolves, its token distribution progressively diverges from the ground-truth distribution on which the decoder was trained, such that reward scores improve while decoded image quality degrades. To address this mismatch, we propose **RankE**, the first end-to-end post-training framework for discrete T2I generation. Rather than optimizing the policy against a fixed decoder, RankE co-evolves the policy and decoder through alternating optimization: the policy is refined via group-relative preference optimization, while the decoder is jointly adapted via reward-aware adversarial training. This co-evolution ends the fidelity--alignment trade-off that plagues frozen-decoder approaches: on LlamaGen-XL (775M), standard RL improves CLIP but degrades FID, whereas RankE simultaneously improves both (FID 15.21, CLIP 33.76 on MS-COCO 30K). Consistent joint gains on Janus-Pro (1B) confirm that decoder co-evolution reliably converts reward optimization into pixel-space quality improvements.

Deep learning models trained to optimize average accuracy often exhibit systematic failures on particular subpopulations. In real-world settings like healthcare, the subpopulations most affected by such disparities are frequently unlabeled, partially observed, or not known in advance. Existing group-robust methods typically assume prior knowledge of the relevant subgroups, using group annotations for training, validation, or model selection. We propose Misclassification Aware Rank-Limited Adaptation (MARLA), a parameter-efficient method for improving worst group performance without explicit subgroup annotations. MARLA leverages an ERM-trained model by calculating the model's misclassification probability scores on a held-out adaptation set to identify a low-dimensional subspace where errors concentrate. We then learn a rank-restricted additive correction to the classifier logits within that subspace. Across seven real-world datasets, we evaluate group robustness under three settings: no knowledge of subgroup relevance, partial knowledge of subgroup relevance, and full knowledge of subgroup relevance. MARLA improves worst-group performance while remaining fast, parameter-efficient, and practical to tune.

Graph representation learning has become a standard approach for analyzing networked data, with latent embeddings widely used for link prediction, community detection, and related tasks. Yet a basic design choice, the latent dimension, is still treated as a brittle hyperparameter, fixed before training and tuned by held-out performance. Learned factors are also identifiable only up to rotation and rescaling, so the nominal rank rarely coincides with the quantity that governs model behavior. We propose Spectral Prefix Extraction and Capacity-Targeted Representation Analysis (Spectra), which replaces rank as the unit of analysis with the spectrum of a learned positive semidefinite kernel, trace-normalized so that spectra are comparable across fits. The normalized eigenvalues form a distribution on the simplex, and their Shannon effective rank acts both as a summary of learned capacity and as a controllable training-time coordinate: a single scalar shapes this realized dimension during training, and bisection targets any desired value within the rank cap. To theoretically support that, we show local regularity and monotonicity of the realized-dimension profile. Across collaboration, social, biological, and infrastructure networks, Spectra traces performance--capacity frontiers that make the trade-off between predictive accuracy and realized dimension visible. It performs competitively with strong link-prediction baselines, yields aligned lower-capacity views of the same fitted model through spectral prefixes, and provides a principled handle on capacity in the overparameterized regime. Capacity thus becomes a property of the fitted model rather than a hyperparameter of the training.

We consider the optimization problem of an expensive-to-evaluate black-box function, in which we can obtain noisy function values in parallel. For this problem, parallel Bayesian optimization (PBO) is a promising approach, which aims to optimize with fewer function evaluations by selecting a diverse input set for parallel evaluation. However, existing PBO methods suffer from poor practical performance or lack theoretical guarantees. In this study, we propose a PBO method, called randomized kriging believer (KB), based on a well-known KB heuristic and inheriting the advantages of the original KB: low computational complexity, a simple implementation, versatility across various BO methods, and applicability to asynchronous parallelization. Furthermore, we show that our randomized KB achieves Bayesian expected regret guarantees. We demonstrate the effectiveness of the proposed method through experiments, including those on real-data emulators.

Vanilla SVGD is known to have a computational cost of $\mathcal{O}(N^2)$ per iteration. In this work, we introduce Random-Projection Tree Stein Variational Gradient Descent (RP-SVGD) to alleviate this computational challenge. This is achieved by restricting kernel interactions to spatially proximal particles clustered via a random projection tree, further reducing to a cost of $\mathcal{O}(N d(\log (N/C)+C))$, where $C$ is a hyperparameter representing the maximum leaf node capacity of the spanning tree. To establish theoretical validity, we introduce a smoothed RP-Tree kernel for any fixed tree realization and prove that it belongs to the Stein class of the target distribution, thereby ensuring the generation of valid gradient flows. In addition, by considering the effective kernel as the expectation of random tree partitions, we verify that it preserves regularity along with other necessary conditions, which guarantee convergence under the approximate gradient flow framework. Extensive experimental results demonstrate that RP-SVGD tends to have competitive performance and significant speedups across various tasks.

Domain generalization methods train on multiple source environments but usually deploy a single pooled predictor. We study a different deployment object: one head per source environment on a shared representation, combined by fixed weights when the test environment label is unavailable. Under a same-meta random-effects model, where environment-specific population minimizers are iid deviations around a shared meta-mean, the optimal fixed simplex rule is the uniform centroid. The same analysis separates this centroid from pooled ERM: under a common feature second moment, pooled ERM is sample-weighted, while the centroid is environment-weighted, giving a closed-form gap controlled by sample-weight imbalance. We introduce **REC** (Random-Effects Centroid), a training objective that fits source-specific experts while directly scoring their averaged deployment head. In the common-second-moment case, the centroid term controls the trained-centroid discrepancy and the dispersion of environment-specific optima. Synthetic experiments isolate the population estimator effect and learned-representation behavior; WILDS experiments show that REC is competitive with established DG baselines and that the centroid term is necessary for a useful averaged predictor.