NeurIPS 2026 Papers

Multimodal LLMs often produce fluent yet unreliable reasoning, exhibiting weak step-to-step coherence and insufficient visual grounding, largely because existing alignment approaches supervise only the final answer while ignoring the reliability of the intermediate reasoning process. We introduce SR-MCR, a lightweight and label-free framework that aligns reasoning by exploiting intrinsic process signals derived directly from model outputs. Five self-referential cues—semantic alignment, lexical fidelity, non-redundancy, visual grounding, and step consistency—are integrated into a normalized, reliability-weighted reward that provides fine-grained process-level guidance. A critic-free GRPO objective, enhanced with a confidence-aware cooling mechanism, further stabilizes training and suppresses trivial or overly confident generations. Built on Qwen2.5-VL, SR-MCR improves both answer accuracy and reasoning coherence across a broad set of visual benchmarks; among open-source models of comparable size, SR-MCR-7B achieves state-of-the-art performance with an average accuracy of 81.4%. Ablation studies confirm the independent contributions of each reward term and the cooling module.

On-policy distillation (OPD) has become a popular training paradigm in the LLM community. This paradigm selects a larger model as the teacher to provide dense, fine-grained signals for each sampled trajectory, in contrast to reinforcement learning with verifiable rewards (RLVR), which only obtains sparse signals from verifiable outcomes in the environment. Recently, the community has explored on-policy self-distillation (OPSD), where the same model serves as both teacher and student, with the teacher receiving additional privileged information such as reference answers to enable self-evolution. This paper demonstrates that learning signals solely derived from the privileged teacher result in severe information leakage and unstable long-term training. Accordingly, we identify the optimal niche for self-distillation and propose \textbf{RLSD} (\textbf{RL}VR with \textbf{S}elf-\textbf{D}istillation). Specifically, we leverage self-distillation to obtain token-level policy differences for determining fine-grained update magnitudes, while continuing to use RLVR to derive reliable update directions from environmental feedback (e.g., response correctness). This enables RLSD to simultaneously harness the strengths of both RLVR and OPSD, achieving a higher convergence ceiling and superior training stability.

Selective state-space models such as Mamba replace fixed linear state transitions by input-dependent transitions, yielding recurrent sequence models that retain linear-time inference while adapting their memory dynamics to the current token. We show that even a simple form of this selective memory update has universal approximation power. Specifically, for every compact input domain, finite horizon, and continuous causal sequence-to-sequence map, a clocked diagonal selective state-space scan with strictly contractive gates and a linear-in-state readout uniformly approximates the target map. The proof is constructive: input-dependent diagonal propagation generates tensor-product partition features over prefixes, and the linear-in-state readout combines these features with sampled target values. Thus the high-order interactions required for universality are produced inside the selective scan itself, rather than by attention, nonlinear hidden-state dynamics, or a universal multilayer perceptron applied after a sequence encoder. These results identify selectivity as a mathematically sufficient mechanism for expressive linear-time state-space sequence modeling. Selectivity lets the model decide, token by token, which parts of the past should be preserved, faded, or combined with the present. The universal approximation theorem shows that this adaptive use of memory is sufficient to reproduce any well-behaved causal sequence rule over a fixed horizon.

Diffusion-based video generation has achieved impressive visual fidelity, yet enforcing temporally consistent motion under explicit control signals remains challenging. Recent noise-warping approaches construct temporally correlated noise by transporting the initial latent noise via diffeomorphic deformation fields derived from optical flow, but this spatial resampling inherently destroys the i.i.d. Gaussian structure of the noise, requiring post-hoc correction steps at the cost of non-trivial computational overhead. We propose \textbf{Seed-Your-Motion}, by conditioning the a priori latent noise on optical flow via per-location orthogonal transformations in latent channel space, parameterized as products of the Householder Transformation. This channel-space formulation has an advantage of preserving the spatial Gaussianity exactly by construction, requiring neither Jacobian corrections nor post-hoc distribution restoration, while at the same time inducing structured cross-frame correlations that ensures geometric consistency across time. With no learnable parameters and a single forward pass, our method is agnostic to model architecture and can be plugged seamlessly into any video diffusion fine-tuning pipeline. Extensive experiments on motion-transfer and camera-controlled video generation benchmarks demonstrate that our method matches or outperform the state-of-the-art noise-warping baseline, while achieving a 2.7× speedup in processing time. The results suggest that the channel-space Householder transformations provide a principled and efficient alternative to noise-warping based motion-controllable video diffusions.

Recent 3D large multimodal models (3D-LMMs) rely on a visual bottleneck to compress complex 3D scene evidence into a limited number of visual tokens compatible with large language models (LLMs). Current visual bottlenecks, however, often passively compress heterogeneous 3D evidence into a homogeneous object-centric token sequence, leaving the spatial organization of the scene under-represented. This under-representation forces the LLM to recover spatial relations from a flattened token sequence, leading to unstable reasoning in relation-intensive and spatially ambiguous scenes. To address this issue, we propose **SceneScaffold**, an active scene-state construction framework for unified 3D scene understanding. SceneScaffold reformulates the visual bottleneck from a passive feature compressor into an active scene organizer, constructing a role-aware spatial scaffold before language reasoning. Specifically, SceneScaffold organizes superpoint-level visual evidence into scene-state components with distinct structural roles: entity states preserve core object semantics, scene-frame states maintain spatial references via boundary and region anchors, relation states encode object-environment interaction cues, and a global summary provides compact context. Through this role-aware construction, SceneScaffold provides the LLM with a spatially organized scene representation before language reasoning. Experiments on unified 3D scene understanding tasks, including 3D visual grounding, question answering, and dense captioning, demonstrate the effectiveness of SceneScaffold, while diagnostic results further show its applicability to relation-intensive and spatially ambiguous cases.

Progress in genomic language modeling depends on pretraining resources that are both diverse and comprehensive, as well as benchmarks that evaluate models across biological scales from genes to whole genomes. This need is especially acute in bacteria, which are extremely abundant, diverse, and central to human health, ecology, and biotechnology. Yet resources for large-scale pretraining and evaluation in bacterial genomics remain limited, with existing frameworks focusing largely on eukaryotic settings. Here, we introduce BacCorpus, a genome-level resource for training genomic LMs comprising 20B proteins, including 4.6B unique sequences, from 7M genomes spanning more than 150,000 species, and BacBench, a comprehensive benchmark spanning 10 datasets and 5 tasks across gene-, system-, and genome-level tasks. We evaluate a diverse suite of genomic LMs on BacBench and additionally develop BacLM, a 350M-parameter model trained jointly on protein sequences and intergenic DNA to further investigate how bacterial genomes are best represented. Across tasks, we find that performance depends strongly on biological scale: methods leveraging protein sequence are competitive on gene-level prediction, whereas models that incorporate wider genomic context excel at problems which involve complex multi-gene interactions. Together, BacCorpus and BacBench provide a practical foundation for training and evaluating robust, genome-aware genomic language models in bacteria at an unprecedented scale.

Guard models are the last line of defense between a language model and a harmful output, yet their training objective is surprisingly narrow. Existing guards learn to predict a single verdict token from a conversational context, concentrating supervision on a single target. The consequences are structural: models latch onto shortcut features, are overconfident, and remain sensitive to where safety evidence appears in the sequence rather than its role in the full context. We propose a different framing. Rather than predicting a label from text, our $\textbf{LLaDA Guard}$ asks which label better explains the text: scoring the prompt or response under each label hypothesis and classifying based on their difference. This shifts supervision to every token in the moderated region, forcing the model to account for full content rather than its most discriminative fragments. We instantiate this idea with a masked diffusion language model, fine-tuning LLaDA-8B-Instruct with a class-conditional reconstruction objective using LoRA and requiring no architectural changes beyond the base model. LLaDA Guard leads on average rank against discriminative baselines trained on stronger backbones across seven held-out safety benchmarks, while exhibiting substantially better confidence calibration (ECE 0.0875 vs. 0.1384 for Qwen3Guard), less over-defense on benign prompts with $\texttt{unsafe}$-looking cues, and less prompt leakage when moderating responses. Its generative nature further enables token-level risk localization as a natural byproduct, yielding a pipeline for rewriting $\texttt{unsafe}$ prompts into $\texttt{safe}$ equivalents without additional training and achieving a 60.7% average conversion-to-$\texttt{safe}$ rate. The weights and code for reproduction are available at the following [anonymized repository](https://anonymous.4open.science/r/LLaDA-Guard2026-B056/).

Frontier AI models and multi-agent systems have led to significant improvements in mathematical reasoning. However, for problems requiring extended, long-horizon reasoning, existing systems continue to suffer from fundamental reliability issues: hallucination accumulation, memory fragmentation, and imbalanced reasoning-tool trade-offs. We introduce **STAR-Math**, a multi-agent framework that systematically addresses these challenges through meta-level supervision and structured Reasoner-Verifier interaction. STAR-Math is structured as an orchestrated state machine with nested challenge-step-replan loops, governed by a reasoning-free Python orchestrator that separates control from inference and bounds error propagation through trace-back and re-planning. Our key innovation is a **persistent Meta-Strategist** that maintains cross-attempt memory and exercises meta-level control by issuing high-level strategic guidance or mandatory directives, so the system can escape unproductive loops rather than stagnate or over-rely on tools. STAR-Math achieves state-of-the-art results on all eight top-tier competition benchmarks: AIME 2025-2026, MathArena Apex Shortlist, MathArena Apex 2025, Putnam 2025, IMO 2025, HMMT February 2026, and USAMO 2026. It obtains perfect scores on AIMEs, Putnam, and HMMT, and shows its largest margin on Apex 2025, scoring 93.75% compared with 80.21% by the strongest baseline GPT-5.5. Ablation studies show that the gains arise from the framework's orchestration rather than from model-level diversity since removing key components or substituting in mixed backbones consistently weakens performance.

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/}.

Queer vernacular is rarely studied in NLP, despite advancements in resources and evaluation for other sociolects and informal language. Because of this, NLP systems often process queer language incorrectly, e.g., they misclassify it as hate speech or generate negative responses. To address this problem, we propose \textsc{Slaying}, the first real-world dataset of English queer slang. \textsc{Slaying} is community-validated, and includes over 400 queer slang terms that pertain to more than 20 queer subcommunities. We argue that queer language data resources have great potential in NLP -- e.g., as components of large pretraining corpora and as the basis for benchmarks -- and can improve queer users' experience of NLP systems. We leverage \textsc{Slaying} for two novel findings in support of this argument: (i) For a number of language models, we show that they are unbiased towards the queer community, but at the same time unable to process its language, i.e., absence of representation bias does not entail the absence of linguistic bias. (ii) Model performance on queer slang varies across queer subcommunities; it is generally worse for slang pertaining to African-American and Latine communities. These findings are relevant for both the queer NLP and the broader ML communities. \textsc{Slaying} is available to the public, and open to future revisions and extensions. \textbf{Warning:} This paper contains profane and potentially offensive language.

Text-driven image editing has advanced rapidly, but reliably localizing these manipulations requires image manipulation localization (IML) models trained on large pixel-annotated datasets, and there is still no low-cost way to obtain such training data at scale. We observe that these data already exist in disguise: public editing datasets contain millions of structurally identical $\textit{(original, edited)}$ pairs to IML training samples, lacking only pixel-level masks. Recovering these masks automatically is non-trivial: pixel differencing is overwhelmed by diffusion-induced perturbations across all pixels, and instruction-only grounding localizes only what the prompt describes, missing unintended editor side-effects. We propose $\textbf{SIGMA}$ ($\textbf{S}$emantic-difference $\textbf{I}$nstruction-$\textbf{G}$rounding $\textbf{M}$ask $\textbf{A}$nnotator), which performs semantic-feature differencing in a vision foundation backbone and injects an instruction-derived spatial prior into this visual stream via bidirectional cross-modal refinement, amplifying the difference signal at intended-edit regions when the editor faithfully realizes user intent. SIGMA is trained in two complementary stages: Stage I supervises on inpainting masks; Stage II closes the diffusion-domain shift via VAE-roundtrip noise calibration, EMA self-training, and an edit-noise disentanglement loss. SIGMA outperforms existing automatic mask generators on five benchmarks ($\textbf{+12.20\%}$ F1, $\textbf{+11.16\%}$ IoU). When applied to public editing corpora, it produces a $\sim$1.1M IML training set that improves six diverse detectors by $\textbf{+18.34\%}$ F1 across five datasets, turning previously unused editing data into a model-agnostic supervisory resource for IML. For reproducibility, our anonymized code is available at https://anonymous.4open.science/r/SIGMA-14FE.

We consider the conversion of musical recordings into human-readable sheet music annotated with timestamps. Such output lets a listener clearly visualize *rubato* (temporally expressive playing), a learner diagnose ensemble precision and timing choices against the written music, and a musicology scholar compare performance styles across recordings of the same work. We introduce (1) a prompt-conditioned encoder--decoder model, named Rubato, trained to output (2) a new textual representation for polyphonic music, named InterMo, which we designed for compatibility with sequence-to-sequence training. Our experiments demonstrate that Rubato produces timestamped piano sheet music from audio with higher notational accuracy than the best existing approaches, which are based on cascades. We find that even if the cascade is given ground-truth MIDI instead of audio, Rubato performs better, suggesting that the ceiling of existing approaches is primarily representational, not acoustic. Further, because Rubato is trained on several related tasks (with prompts), it competes with or outperforms the best single-task systems on related but simpler tasks like MIDI note grounding and beat/downbeat detection. We will release the model and code on publication, and an anonymized demo is available at https://rubato-transcription.github.io.

Irregular multivariate time series (IMTS) are sampled at non-uniform timestamps and asynchronous across channels, their dependencies are shaped by time gaps, local ordering, and channel-specific sampling patterns. Building a faithful attention mechanism for IMTS exposes two limitations of existing work. First, current irregular-time attention methods typically regularize raw inputs into interpolated, aligned, or continuous-time-embedded forms before applying attention, which blurs exactly these irregular cues that determine how two observations relate. Second, beyond representation, standard dot-product attention scores pairs through learned token projections with no intrinsic tie to the shape of the underlying series, and an attention that does reflect this shape becomes costly to compute in high-channel IMTS. We propose Rubato, which reads two views directly from the raw irregular input: per-channel local dynamics, and a path signature that summarizes the joint multi-channel series. Attention in Rubato then scores pairs by a signature-kernel measure of series similarity rather than by a generic dot product, and a structured random projection (SORF) keeps this comparison affordable when the channel count is large. We prove that the structured projection introduces only a controlled bias that vanishes as the number of channels grows, and experiments on multiple IMTS benchmarks show that Rubato consistently outperforms strong baselines.

Standard ensemble methods can improve robustness by mitigating individual model stochasticity, but they frequently fail to generalize beyond the training distribution, leaving systems vulnerable to out-of-distribution (OOD) data. We introduce the Robust Satisficing Ensemble (RSE), a scalable probabilistic aggregation framework designed to secure model combinations by minimizing their fragility against distributional changes. By bridging a robust satisficing objective with a quadratic risk surrogate, RSE derives a closed-form analytical inner adversary. This fundamental reformulation enables highly efficient weight updates without the computational bottleneck of standard minimax optimization, crucially decoupling optimization complexity from sample size to scale effectively to large datasets. In particular, we prove a geometrically decaying truncation error for the fragility approximation via Lagrange-Bürmann expansion and show that the proximal RSE updates converge to a stationary point. We evaluate RSE across a comprehensive spectrum of distribution shifts, including adversarial label shift (SST-5, TREC), natural subpopulation and temporal shifts (CivilComments, HuffPost). Our results demonstrate that RSE consistently outperforms strong robust baselines, including GroupDRO, IRM, and SRM. These findings establish RSE as a practical, theoretically grounded safeguard for reliable model aggregation in dynamic environments.

Fine-tuning large language models (LLMs) on downstream tasks often weakens their previously aligned safety behavior, creating an inherent trade-off between task adaptation and safety preservation. Existing approaches attempt to mitigate this issue by identifying safety-relevant neurons, layers, or update directions in the original parameter space. However, safety-related information is often distributed across many directions in this space, causing downstream updates to overwrite parameters that encode safety behavior. We propose WSR-Tune, a Weight Space Reparameterization-based fine-tuning framework that preserves safety during downstream adaptation. Our approach constructs a safety-conditioned basis and reparameterizes the weight matrices such that safety-relevant information becomes concentrated in a small subset of directions. We then identify safety-critical directions and freeze them in the reparameterized space, while updating the remaining complementary directions for downstream adaptation. By explicitly structuring the parameter space to disentangle safety and task information, our method reduces interference between objectives and enables simultaneous preservation of safety and improvement in downstream performance. Experimental results show that our method achieves a more favorable trade-off between downstream performance and safety retention, demonstrating its effectiveness for reliable LLM fine-tuning.

Upgrading a deployed face-recognition model usually makes new query embeddings incompatible with the legacy gallery, forcing costly backfilling of stored embeddings. We show that this incompatibility is often much simpler than it appears: independently trained angular-margin face models exhibit representation rigidity, where their embedding spaces are well approximated by a single global orthogonal transformation. We formalize this phenomenon as Procrustes rigidity, derive a finite-calibration recoverability bound, and estimate the orthogonal adapter from a small paired calibration set via closed-form Procrustes alignment. The resulting deployment protocol is backfill-free: transport upgraded queries into the legacy coordinate system, then reuse the existing gallery without re-embedding stored identities. Across 45 models spanning three architectures, two losses, three data scales, and five seeds, we find strong rigidity within architecture families, non-monotonic scaling with data size, and robust cross-model retrieval. On the 1.6M-image IFRT benchmark, Procrustes-aligned queries recover at least $99.8\%$ of mono-model Rank-1 accuracy for same-family upgrades, while lightweight residual correction substantially narrows the cross-architecture gap.

To date, the field of reinforcement learning (RL) has primarily focused on developing discounted methods, which are designed to optimize the short-term or transient performance of RL agents. Conversely, the field has been reluctant to explore and invest in methods that are designed to optimize the long-term or steady-state performance of RL agents. In this position paper, we challenge this reluctance under the framing that RL agents should be viewed as competitive swimmers. That is, we argue that, like competitive swimmers, RL agents cannot hope to win all races, or solve all tasks, based on transient performance alone. To this end, we first motivate the need for RL agents that can adequately optimize both the transient and steady-state performance. Then, we summarize recent works which show that discounted methods have deficiencies when it comes to optimizing the steady-state performance. Finally, we argue in favor of more attention and investment from the community towards average-reward methods, which are designed to optimize the steady-state performance. Namely, we challenge common misconceptions associated with average-reward methods, and highlight recent advancements which suggest that such methods present a viable and potentially superior alternative.

Fast weight architectures offer a promising alternative to standard transformers for long-context modeling by replacing the KV-cache with a recurrently updated fixed-size memory. Despite this architectural shift, they are typically trained with the same next-token prediction (NTP) objective as standard transformers. This creates a mismatch: fast weight models rely on an evolving memory state to support future predictions, while NTP provides only token-level supervision for the immediate next token. We address this mismatch with next-sequence prediction (NSP), a sequence-level extension of NTP that trains models to produce coherent multi-token continuations from their recurrent memory state. To optimize this sequence-level objective, we propose ReFINE ($\textbf{Re}$inforced $\textbf{F}$ast we$\textbf{I}$ghts via $\textbf{N}$ext s$\textbf{E}$quence prediction). ReFINE selects informative token positions based on prediction entropy, generates multi-token rollouts, assigns sequence-level rewards, and optimizes the model with Group Relative Policy Optimization (GRPO). ReFINE is applicable throughout the training lifecycle of pre-trained language models: mid-training, post-training, and test-time training. Experiments on LaCT-760M and DeltaNet-1.3B show that ReFINE consistently outperforms NTP-based supervised fine-tuning in needle-in-a-haystack retrieval, long-context question answering, and diverse tasks in LongBench. ReFINE provides an effective and versatile framework for improving long-context modeling in fast weight architectures.

Agnostic online learning is classically solved via a reduction to the realizable setting, utilizing Littlestone's Standard Optimal Algorithm (SOA) as a base learner. However, the SOA is computationally intractable to execute even for a single round. To overcome this barrier, recent work in oracle-efficient online learning replaces the SOA with a realizable base learner that accesses the concept class exclusively through an offline empirical risk minimization (ERM) oracle. While such agnostic learners achieve near-optimal expected regret, they suffer from a doubly-exponential oracle complexity of $\mathcal{O}\big(T^{2^{\mathcal{O}(d_\mathrm{LD})}}\big)$, where $d_\mathrm{LD}$ is the Littlestone dimension and $T$ is the number of rounds. In this work, we significantly improve this oracle complexity while relying on an even weaker primitive: a weak-consistency oracle, which merely decides whether a given labeled dataset is realizable. At the core of our approach is an adaptive and dynamic agnostic-to-realizable reduction that actively prunes non-realizable label sequences on the fly. By using the VC dimension ($d_\mathrm{VC}$) to bound the number of dynamically maintained active paths, our algorithm reduces the total query complexity down to $\mathcal{O}(T^{d_\mathrm{VC}+1})$ while perfectly preserving near-optimal expected regret. Crucially, this dynamic pruning also yields a memory reduction over the standard reduction. Furthermore, we formally quantify the regret--oracle complexity tradeoff, providing upper bounds that smoothly interpolate between restricted query budgets and attainable expected regret. We complement these with lower bounds proving that any learner restricted to $Q = o(\sqrt{T})$ queries must suffer an expected regret of $\Omega(T/Q)$.