NeurIPS 2026 Papers

Chunked prefill has become a widely adopted serving strategy for long-context large language models, but efficient attention computation in this regime remains challenging. Existing sparse attention methods are primarily designed for one-shot prefill and do not translate efficiently to chunked prefill: block-sparse kernels lose efficiency when the query length is limited by the chunk size, while fine-grained pattern search becomes costly when repeated over the accumulated KV cache at every chunk. QUOKA, a recent method that directly targets chunked prefill, avoids sparse-kernel overhead but relies on query-subsampled, token-level KV selection, which can miss query-specific KV entries and introduce explicit KV-copy overhead. To address these limitations, we propose CompactAttention, a chunked-prefill attention mechanism based on Block-Union KV Selection. CompactAttention treats 2D block-sparse masks as KV-selection signals rather than direct sparse-kernel execution plans, and converts them into GQA-aware per-group KV block tables through Q-block union and intra-group union. This construction produces the minimal block tables that preserve all KV blocks selected by the input masks under paged execution constraints, enabling selected KV blocks to be accessed in place without explicit KV compaction. On LLaMA-3.1-8B-Instruct, CompactAttention maintains accuracy close to dense attention on the RULER benchmark while delivering up to 2.72$\times$ attention speedup at 128K context length under chunked prefill.

AI systems are increasingly used in community contexts, yet they often fail to align with the communities they impact. To address this misalignment, we advocate for community-centered AI---an AI development paradigm that shifts away from pursuing general-purpose AI and instead focuses on co-developing tailored AI systems with the specific communities they impact. This position paper argues that community-centered AI is both feasible and beneficial for impacted communities. To support this argument, we first outline the potential benefits of community-centered AI, including greater alignment, enhanced legitimacy, capacity building, and normative justification. Next, we highlight emerging efforts in community-centered AI across the AI development process, demonstrating its feasibility in delivering these benefits. Finally, we propose a framework to guide future work in this area, characterizing community-centered AI efforts along three dimensions: participation mode, decision-making process, and development layer. Overall, this paper calls for increased research on community-centered AI.

The Hessian of the energy with respect to the nuclear positions is indispensable in atomistic modelling. However, constructing this matrix requires $O(N)$ Hessian vector products, traditionally limiting high-accuracy Hessians to small systems. Machine learning interatomic potentials (MLIPs) have accelerated atomistic modelling by providing highly accurate energies and forces at $O(N)$ cost, yet the resulting $O(N^2)$ cost of Hessians remains a practical bottleneck for large systems. Based on the insight that we can derive the sparsity pattern for an MLIP's Hessians in closed form, we show in this paper how to use techniques from sparse automatic differentiation to reduce the cost of a local MLIP's Hessians to a system-size-independent number of Hessian-vector products, yielding overall $O(N)$ total cost without any approximations. We benchmark our approach on a variety of systems ranging from alkane chains to water clusters to A$\beta$40 conformers. Depending on the MLIP configuration, we achieve the linear scaling regime already on relatively small systems, resulting in large runtime reductions between $2\times-15\times$ for these systems. This opens up the possibility of scaling high-accuracy MLIP Hessians to very large systems, such as proteins that were previously inaccessible.

Data scaling is fundamental to modern deep learning, and grows increasingly critical as autonomous driving shifts to end-to-end learning. Real-world driving data is expensive to annotate and scene-biased, making real-synthetic co-training with near-infinite synthetic data a promising direction. However, naively incorporating all available synthetic data is inefficient and leads to distribution shifts, and optimizing data mixture under practical training budgets remains a critical yet under-explored problem. In this sense, we claim that the mixture of training data requires clear guidance in terms of scene types and quantities. Particularly in this work, we conceptualize the data mixture approximately as a dynamic optimization process that iteratively adjusts the training data mixture to maximize model performance, guided by closed-loop evaluation feedback, and propose AutoScale, a fully automated closed-loop data engine unifying scene representation, data mixture optimization and retrieval, as well as model training and evaluation. Specifically, we propose Graph Regularized AutoEncoder (Graph-RAE) for driving scene representations, introduce Cluster-aware Gradient Ascent (Cluster-GA) for cluster-wise importance estimation and reweighting, and perform cluster-guided vector retrieval to select high-value samples. Experiments on NavSim demonstrate that AutoScale outperforms vanilla co-training and cross-domain baselines, achieving better performance with fewer synthetic samples under constrained budgets.

Many real-world problems demand exact constraint satisfaction to be meaningfully solved, yet existing neurosymbolic methods face a stark trade-off: those that guarantee satisfaction fail to scale, while those that scale must resort to approximation. Meanwhile, traditional deep learning approaches offer no satisfaction guarantees at all. In this paper, we propose Casper, a neurosymbolic layer that simultaneously (i) guarantees constraint satisfaction, (ii) scales at inference time, (iii) is fully automated, requiring no problem-specific engineering, and (iv) is fully differentiable, making it composable with any neural predictor both at training and at inference time. Casper casts constraint satisfaction as a closed-form Euclidean projection onto the constraint-feasible region, computable in a single forward pass. Experiments across MNIST arithmetic (up to 1024 digits), Sudoku solving, and Warcraft pathfinding show that Casper is the only method that produces guaranteed-valid outputs at scale: on 30×30 Warcraft grids, exact baselines time out and approximate ones produce valid paths less than 7% of the time, while Casper produces them 100% of the time.

As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual signals becomes a safety imperative. Existing mechanistic analyses remain confined to unimodal text and offer no explanation for why a single misleading sentence can override a correct image based diagnosis, or why a model commits to a confident answer despite insufficient visual evidence. We find that these two safety risks, arbitration failure where textual context overrides visual grounding and brake failure where the model commits without adequate evidence, are mediated by spatially disjoint attention head populations: arbitration heads form a mid-to-deep wideband reflecting cross-layer evidence competition, while brake heads concentrate in a narrow middle-to-late layer band that regulates evidence sufficiency and abstention behavior. To ground these observations in causal circuitry, we introduce CRAFT, which localizes each failure mode to a minimal causal head set via dual criteria and verifies necessity and sufficiency through temporal probes and Tuned Lens trajectory analysis. Excising arbitration heads sharply reduces conflict following with negligible degradation on clean inputs, while excising brake heads restores appropriate abstention under degraded visual evidence. The two interventions target spatially disjoint head sets and produce distinct corrective effects, underscoring the mechanistic separability of the failure modes. Experiments across multiple medical VQA benchmarks and VLM architectures validate both the localization and interventions, demonstrating that the identified heads causally drive each failure mode and that targeted modulation generalises without retraining. The code is available at \url{https://anonymous.4open.science/r/CRAFT-B3CA}.

This paper studies the problem of semi-supervised edge classification, which aims to identify edge relations using both labeled and unlabeled data. This problem is highly challenging due to the inherent asymmetry of edge relations and strong prediction biases from imbalanced homophilous and heterophilous edges. Towards this end, this paper proposes a novel approach named Complementary Tripartite Play with Bayesian CaLibration (CLEAR) for semi-supervised edge classification. The core of our CLEAR is to incorporate asymmetric semantic branches and a meta-teacher into a tripartite-play framework, enabling reliable guidance and optimization under label scarcity. In particular, our CLEAR introduces a topological branch and a contextual branch which extracts graph semantics, generating pseudo-labels in complementary views. To improve the reliability of pseudo-labels, we filter low-quality pseudo-labels using epistemic uncertainty, and then calibrate posterior distributions using the priors induced by neighborhood information. More importantly, we utilize a meta-consoler to guide the optimization of two branches by outputting the objective coefficients, ensuring a reliable pseudo-labeling process in a tripartite play framework. Extensive experiments on benchmark datasets validate the superiority of the proposed CLEAR in comparison to various baselines. Our code is available at https://anonymous.4open.science/r/CLEAR-D1D7/.

Hallucination detection is critical for deploying large language models (LLMs) in real-world applications. Due to strong empirical performance, internal representation–based methods have emerged as the prevailing direction for detecting hallucinations, yet they remain largely model-specific and often fail to generalize to unseen LLMs. In this paper, we study an important yet underexplored problem, termed cross-model hallucination detection (CMHD), which aims to train hallucination detectors on source LLMs while ensuring performance on unseen target LLMs. The core challenge of CMHD lies in cross-model representation heterogeneity: hidden states from different LLMs exhibit severe semantic inconsistency, which limits the transferability of classical representation-based detectors. Through analysis and empirical validation, we show that relational structures over layers, tokens, and features capture transferable detection signals despite semantic inconsistency. Based on this observation, we propose cross-model hallucination detection via heterogeneity-oriented relational discrimination (CHORD), a relational graph discrimination framework that represents hidden states as joint relational graphs over layers, tokens, and features. It feeds these graphs into relation-aware attention to obtain structure-aware features, and uses meta-learning to shape these features toward transferable detection signals. Extensive experiments show that CHORD outperforms representative detection methods in cross-model generalization.

Street-level bureaucrats, such as caseworkers, triage nurses, and school administration, operate under a fundamental tension. The official policy on who to prioritize for services may conflict with bureaucrats estimation of who would benefit most from those services. The social theory of street-level bureaucracy argues that frontline agents use discretionary authority to bridge this gap, overriding policy recommendations when professional judgment identifies cases where doing so would meaningfully improve outcomes. As algorithmic recommendations become embedded in allocation settings, understanding the structure of this discretionary behavior is fundamental to designing tools that account for, rather than suppress, this value of professional judgment. We formalize discretion as a finite-horizon dynamic allocation problem. An agent implements a default policy but holds a limited budget of K overrides across T periods. Each period, the agent observes whether deviating from the default would yield a welfare gain and must decide whether to spend a scarce override now or conserve it for future opportunities. We show that the optimal policy is a state-dependent threshold rule and prove a behavioral invariance: for location-scale families of improvement distributions, the rate at which an optimal agent exercises discretion is independent of the amount of welfare gains and depends only on the distribution's shape. Fat-tailed gains induce patience (conserving overrides for rare, high-stakes cases); thin-tailed gains induce routine, front-loaded spending. We also compare the conceptual implications of our model with operational data from a homelessness service system. Override rates shift with short-run capacity openings in scarce housing, rising when recent exits free slots, and falling when interventions are congested. This could be the result of a perceived increase in the discretionary budget. Further, the temporal pattern of spending is front-loaded: override rates peak on Mondays and early in the fiscal year, consistent with the routine, early-spending behavior our model predicts under thin-tailed improvement distributions. These findings identify a systematic structure in both human and theoretically optimal decision-making, with direct implications for the design of human-AI allocation systems that are both more efficient and equitable.

Mixture-of-Experts (MoE), popularized by large language models, is a promising paradigm for scaling visual generative models. However, conventional token-wise MoE routes tokens independently within a homogeneous expert pool and regularizes expert usage toward uniformity, making it poorly matched to video data that is spatiotemporally redundant and semantically long-tailed. We show that existing visual MoEs fall into a uniformity trap: semantically under-organized routing, compounded by uniform expert-usage regularization, scatters coherent patches across disparate experts, causing routing fragmentation and structural distortion. To address this, we propose SplitMoE, a split-role sparse architecture that breaks the shackles of uniformity. To accommodate the inherent semantic imbalance, we explicitly bifurcate the expert pool into semantic experts and generic experts, with semantic experts capturing high-level semantic abstraction and generic experts preserving residual visual information and flexible generative capacity. Leveraging prototype-guided routing and push-pull regularization, SplitMoE enables tokens to cluster naturally by semantic attributes rather than arbitrary balancing constraints. Extensive results show that under an equivalent activated-parameter budget, SplitMoE outperforms traditional load-balanced MoEs in convergence speed, routing coherence, and video generation quality across standard benchmarks. By revealing an emergent coarse-to-fine denoising logic, SplitMoE provides the community with a modality-aware scaling path, serving as a critical reference for building large-scale video world models.

Sparse tuning is widely used to adapt large language models due to its parameter efficiency. However, its effectiveness depends not only on the number of trainable parameters, but also on the feature-interaction topology induced by the sparse update. Limited or uneven cross-dimensional interactions may isolate subsets of feature dimensions, forming information islands that hinder task-specific information mixing. To address this issue, we propose **S**mall-**W**orld **I**nduced **F**ine-**T**uning (**SWIFT**), a sparse tuning framework that induces small-world connectivity in parameter space. SWIFT combines block-diagonal sparse updates for local feature interactions with a fixed sufficiently scattering permutation that routes these updates to distant feature groups, creating sparse long-range shortcuts without additional trainable parameters. We further show theoretically that random or block-diagonal sparse tuning can suffer from information islands, whereas SWIFT restores positive expansion through permutation-induced routing. Extensive experiments on commonsense reasoning, natural language generation, and image classification demonstrate that SWIFT consistently outperforms competitive PEFT baselines and can match or surpass full fine-tuning.

Graph contrastive learning constructs augmented views and aligns corresponding instances across views to learn view-invariant representations. However, existing methods usually enforce cross-view consistency within a single representation space, lacking an explicit characterization of view-specific information. This may suppress task-relevant complementary information, reduce embedding diversity, and induce dimensional collapse. To this end, we propose Manifold-Guided Representation Disentanglement (MGRD), a plug-and-play framework for boosting graph contrastive learning. MGRD decomposes graph representations into shared and complementary subspaces. The shared subspace captures view-invariant semantics while preserving local graph structure through a graph-anchored manifold prior, whereas the complementary subspace captures residual view-specific information beyond the shared representation and is regularized by asymmetric consistency and subspace decorrelation. By jointly exploiting shared and complementary representations, MGRD balances cross-view invariance and representation diversity. Experiments on node-level and graph-level benchmarks show that MGRD consistently improves multiple GCL backbones, while theoretical and empirical analyses demonstrate its effectiveness in mitigating dimensional collapse.

Contrastive vision-language (VL) pretraining in computational pathology has generally relied on curated image-caption datasets sourced from social media, educational videos, and research publications. These corpora are sparse relative to the diversity of tissue morphology and difficult to scale further, as pathologists do not routinely produce descriptive captions in clinical practice. We propose to instead leverage slide-level pathology reports which can be generated during routine diagnostic workflows in a scalable manner. A key challenge in this weakly-supervised setting is that pathology reports describe macroscopic, diagnostic findings while patch-level visual features describe microscopic findings. We introduce Locked Image Multiple Instance PreTraining (LIMIT), which addresses this gap by freezing a pretrained vision encoder and fine-tuning only the text encoder: report embeddings serve as cross-attention queries over bags of patch features, producing report-conditioned slide representations optimized via cross-entropy. Using LIMIT, we establish CONCH-Z, which aligns the text encoder of CONCH v1.5 to 335,645 clinical reports. Evaluated across 20 tasks spanning patch-level classification, tumor detection, and slide-level subtyping, CONCH-Z establishes state-of-the-art zero-shot performance over VL encoders trained on curated caption datasets, while simultaneously closing the gap with slide-level foundation models trained with substantially more parameters and compute. We will release CONCH-Z weights and evaluation code to support reproducibility and broader community use.

Modern retrieval-augmented systems typically access a corpus through a fixed similarity interface (either sparse or dense) that returns only the top-$k$ most similar documents for a query. While efficient, this design becomes a bottleneck for agentic search: exact lexical matching, conjunctions of weak clues, local context verification, and multi-step hypothesis refinement are difficult to realize with a conventional off-the-shelf retriever, and evidence filtered out early cannot be recovered downstream. This problem becomes particularly acute when agents must iteratively discover intermediate entities, combine partial evidence, and revise their search strategy. To address this problem, we study direct corpus interaction (DCI), a simple alternative in which an agent searches the raw corpus directly using general-purpose terminal tools (e.g., grep, file reads, shell commands, and lightweight scripts), without relying on embedding models, vector indexes, or retrieval APIs. This approach requires no offline indexing and adapts naturally to evolving local corpora. Across benchmarks for information retrieval (i.e., BRIGHT and BEIR) and end-to-end agentic search (i.e., BrowseComp-Plus), this simple setup substantially outperforms strong sparse, dense, and reranking baselines. It also attains strong accuracy on multi-hop question answering without relying on any conventional semantic retriever. Our results suggest that as language agents grow more capable, retrieval quality depends not only on reasoning ability, but also on how richly the corpus interface supports interaction. DCI therefore points to a broader interface design space for agentic search. Our code is available at https://anonymous.4open.science/r/dci-agent/.

Current multivariate time series forecasting methods mainly rely on static linear decoders, but these often suffer from severe representational bottlenecks. In this paper, we propose a novel architecture called DecodeTS (\textbf{\underline{D}}ynamic \textbf{\underline{E}}xpert-\textbf{\underline{C}}oupled \textbf{\underline{O}}ptimal \textbf{\underline{DE}}coder for Time Series Forecasting), which replaces the conventional static prediction head with a heterogeneous expert library. DecoderTS adopts a divide-and-conquer strategy to disentangle complex temporal dynamics, such as long-term trends and abrupt changes. Crucially, we introduce an optimal transport (OT)-based dynamic routing mechanism that can adaptively assign customized combinations of experts to different variables. By imposing OT marginal constraints, DecoderTS is theoretically proven to achieve collaborative load balancing among experts and effectively eliminate representation collapse. Extensive experiments on more than \textbf{15} datasets and about \textbf{30} baselines demonstrate that DecoderTS achieves state-of-the-art forecasting accuracy, which delivers a \textbf{17.06\%} performance gain while attaining up to about \textbf{41×} inference time speedup and up to \textbf{3.61×} reduction in memory footprint.

Activation steering provides a lightweight inference-time mechanism for controlling large language models (LLMs) by modifying their internal activation vectors toward desired behaviors. Most existing methods compute a fixed steering direction in the original activation space, typically from pairs of contrastive examples using mean differences, linear probes, or arbitrary separability criteria. While effective to a certain extent, these methods treat behavioral control as a global, linear, additive offset: the same direction is applied across inputs, and behaviors are linearly separable. This can be restrictive when behavioral features vary nonlinearly across the activation space or lie on curved and anisotropic manifolds, where the optimal intervention may be input-dependent. To address this limitation, we propose INNSteer, a nonlinear activation steering framework based on invertible latent transformations. Rather than searching for a better steering vector in the original representation space, INNSteer learns a lightweight invertible neural network $\phi$ that maps an LLM's activations into a latent space where behavioral classes are more amenable to linear control. At inference time, activations are mapped through $\phi$, steered in the latent space, and mapped back through the exact inverse transformation $\phi^{-1}$. This makes a simple latent-space translation become a nonlinear, input-dependent intervention in the original activation space. Across experiment settings on multiple LLM families, scales, behavioral traits, and safety benchmarks, INNSteer consistently improves model control over linear, transport-based, and nonlinear steering baselines while largely preserving generation fluency.

Language models (LMs) obtain much of their capability from the diversity of their training data. This is possible because neural architectures like transformers are remarkably flexible and can handle a variety of domains such as natural language, code, and mathematics. This raises the question: are transformers optimal for any specific domain? In this position paper, we present empirical evidence that the inductive biases of transformer-based LLMs should be re-examined, since, while broadly useful, they are often suboptimal on subsets of their training data. **Methods.** To support this position, we use a method that learns new activation functions within a transformer as a tool to explore the space of inductive biases. We train the resulting modified architectures on various data subsets. A large-scale evaluation shows that vanilla transformers are relatively easy to improve, but mostly on a per-domain basis. For example, an architecture optimized for a natural language dataset like FineWeb performs well on another such dataset like TinyStories, whereas architectures optimized for mathematics perform poorly on natural language. We also identify the mechanisms responsible for these improvements. **Implications.** These results call for re-evaluating core design choices in language models. We do not advocate a return to handcrafted, domain-specific models. Instead, we emphasize that different domains can benefit from different learning mechanisms. This will require new designs that preserve cross-domain interactions, analogous to how reasoning in the brain arises from the interaction of multiple specialized networks. We argue that this perspective may be key to improving reliability and data efficiency.

Imitation Learning (IL) is a natural framework for learning in sequential decision-making systems and has emerged as the dominant paradigm through which we understand language model training. A central puzzle is that, while in theory offline IL can be horizon-free and optimal, in practice online methods such as on-policy distillation (OPD) often outperform offline methods such as supervised fine-tuning (SFT). We propose a noisy expert model to explain this gap, in which the learner only has access to a noisy version of the expert's policy, but wishes to compete against the reward achieved by a clean expert, motivated by the fact that in many applications, e.g. training language models to perform long chains of thought, the expert is often imperfect. In this setting, we show a sharp separation between offline and online IL. Offline learning from noisy trajectories is fundamentally hard: to compete with the clean expert, the sample complexity must grow exponentially, in contradistinction to the clean expert setting where no explicit horizon dependence exists. In contrast, we prove that online interaction with the noisy expert via a novel variant of OPD enables horizon-free guarantees in some settings and polynomial dependence on horizon in general. Our analysis leads to an alternative loss function form that is commonly considered empirically for LM training. We further provide algorithms and lower bounds, and extend our results to the more realistic setting of unknown corruption when the clean expert is deterministic, thereby providing a theoretical foundation for why on-policy distillation can outperform standard supervised fine-tuning when training language models from imperfect teachers. We complement our theoretical results with experiments on synthetic and natural-language tasks, showing that the OPD variant suggested by our theory outperforms both offline BC and existing OPD objectives under noisy expert feedback.

Children's earliest years are a period of intense social learning, during which meaning is built through gaze, gesture, and physical contact. Yet, this critical aspect of social interaction remains under-evaluated in multimodal video understanding, as existing benchmarks prioritize the conversation-driven interactions of adults. We address this gap with NEST ($\textbf{N}$aturalistic $\textbf{E}$arly-childhood $\textbf{S}$ocial in$\textbf{T}$eractions), the first child-centric benchmark for nonverbal social reasoning. NEST comprises 1,208 manually annotated clips ($\sim$12s) curated for diversity across naturalistic settings, cultural contexts, and developmental stages. Beyond a lack of data, targeted model improvement in this domain is hindered by coarse evaluation schemas that conflate basic perception with higher-order reasoning. To resolve this, NEST introduces a compositional schema that builds atomic interaction cues into progressive levels of social reasoning: contextual, behavioral, interpersonal, and open-ended descriptions. Extensive analysis of state-of-the-art Vision-Language Models (VLMs) reveals a stark performance collapse: despite strong contextual recognition, models exhibit substantial gaps in decoding precise behavioral and interpersonal cues. We trace these failures to specific biases, including poor consequence grounding and temporal sensitivity. Finally, we ground NEST in real-world applications by evaluating VLMs using domain adaptation strategies. Ultimately, NEST provides a rigorous testbed for advancing socially grounded artificial intelligence in early childhood. The dataset and code will be publicly released.