NeurIPS 2026 论文列表

Modeling hyperspectral imagery (HSI) across different sensors presents a fundamental challenge due to variations in wavelength coverage, band sampling, and channel dimensionality. As a result, models trained under a fixed spectral configuration often fail to generalize to other sensors. Existing Vision Transformer (ViT) approaches either rely on implicit spectral modeling with fixed channel assumptions or adopt explicit spatial–spectral attention with prohibitive computational cost, leading to a fundamental trade-off between efficiency and expressiveness. In this work, we introduce Low-rank Efficient Spatial–Spectral ViT (LESSViT), a sensor-flexible architecture for cross-spectral generalization. LESSViT is built on LESS Attention, a structured low-rank factorization that models joint spatial–spectral interactions through separable spatial and spectral components, reducing the complexity of full spatial–spectral attention from $\mathcal{O}(N^2 C^2)$ to $\mathcal{O}(rNC)$, where $N$ is the number of spatial tokens, $C$ is the number of spectral channels, and $r$ is the rank of the low-rank approximation. We further incorporate channel-agnostic patch embedding and wavelength-aware positional encoding to support flexible spectral inputs. To enable efficient and robust pretraining, we introduce a hyperspectral masked autoencoder (HyperMAE) with decoupled spatial–spectral masking and hierarchical channel sampling. We evaluate LESSViT under a cross-spectral generalization setting that simulates cross-sensor variability. Experiments on the SpectralEarth benchmark demonstrate that LESSViT improves robustness under spectral shifts while remaining competitive in-distribution, and explicit and efficient spatial–spectral modeling is essential for scalable and generalizable hyperspectral representation learning.

We propose LEGO, a budget-aware Dense-to-MoE methodology that converts pretrained vision-language models into efficient FFN-MoE variants under a target architectural budget $((P_{total},P_{active}))$. Our study is motivated by a counter-intuitive recovery pattern: fixed activation ratios that work well for some dense backbones can degrade sharply on others under the same recovery recipe, indicating that raw sparsity alone is an unreliable design rule. Through controlled recovery experiments across VLM families and scales, we identify the activated-FFN-to-hidden ratio $(D_I/D_H)$ as a backbone-comparable structural factor and summarize empirical sizing rules for avoiding severe bottlenecks while limiting diminishing returns. LEGO operationalizes these rules through a discrete budget-aware search over Split, Upcycle, and Hybrid constructions, and introduces a moment-matching scaling factor to reduce the initialization scale shift caused by sparse FFN aggregation. Finally, we use a two-stage multimodal recovery recipe that first stabilizes the sparse LLM with the vision tower frozen, then performs joint fine-tuning to recover the full VLM. Under matched data and training protocols, LEGO improves Dense-to-MoE recovery quality and turns costly blind architecture search into a small set of rule-guided candidates. Codes can be found in supplementary materials.

Despite the remarkable progress of evolutionary LLM frameworks like AlphaEvolve in program and algorithm discovery, they consistently struggle with tasks requiring deep algorithmic reasoning and long-horizon planning, such as the Abstraction and Reasoning Corpus (ARC-AGI). A fundamental limitation of current evolutionary code generation is its reliance on unguided stochasticity to derive child programs from parents, lacking a systematic mechanism for knowledge accumulation, whereas human problem-solving thrives on the continuous consolidation and reuse of high-level conceptual abstractions. To bridge this gap, we propose the \textbf{Library-driven Evolutionary Abstraction Paradigm (LEAP)}, a cognitive-inspired framework that enables LLMs to autonomously construct and evolve a globally shared library of reusable algorithmic primitives. The evolution of this library is governed by two rigorous theoretical principles: First, the generation of new primitives is driven by a Minimum Description Length (MDL) objective, ensuring the system only assimilates concepts that genuinely compress the problem space. Second, the evolutionary survival and selection of these primitives are governed by a principled exploration-exploitation mechanism, dynamically regulating the library's life-cycle to strike an optimal balance between exploiting proven cognitive operators and exploring novel hypotheses while strictly bounding context entropy. Extensive experiments demonstrate that LEAP significantly outperforms existing LLM evolutionary baselines on the ARC-AGI benchmark and math optimizations problems, exhibiting robust cross-task generalization, highly efficient context utilization, and autonomous concept discovery capabilities.

Long-context inference in Large Language Models (LLMs) suffers from high latency because attention computation and memory traffic scale linearly during decoding. KV cache pruning reduces this cost by retaining only a subset of cached tokens, but its effectiveness depends on whether the pruning proxy can recover the tokens with the largest exact attention logits under the current query. Existing proxies suffer from ranking misalignment: static heuristics rely on historical statistics and thus ignore the current query, while hashing-based retrieval proxies rank candidates in a discrete metric space that does not coincide with attention's inner product. To avoid these problems, we propose LAPrune, a KV pruning framework that uses additive vector quantization to build an efficient attention-logit-aligned proxy. LAPrune represents each cached key as a sum of learned codewords and decomposes the query-key dot product into lookup-and-add operations over query-codeword inner products. This replaces high-dimensional dot products with lightweight scalar lookups while preserving fidelity to the original attention ranking. LAPrune requires no base-model finetuning and can be integrated into long-context decoding with minimal modification. Across long-context evaluations, LAPrune improves exact-logit Top-$k$ recovery and perplexity over hashing-based pruning proxies, achieving up to $2.73\times$ decoding speedup over FlashAttention-2 and $4.83\times$ end-to-end speedup over the vanilla baseline at 128K context. Our code is available at [Anonymous/LAPrune](https://anonymous.4open.science/r/LAPrune-VQ).

We present LAION-BVD (LAION - Big Video Dataset), a large-scale open video dataset for multimodal learning, containing _1.3B platform-specific video URLs_ collected from CommonCrawl. From these, we download 80M videos with a total duration of _10 million hours_. The dataset is designed for multimodal pre-training across video, audio, and image modalities. Using content-aware scene detection we extract clips for which we synthetically generate video and audio captions. Models trained on these data achieve competitive performance on standard video-text and audio-text benchmarks, with consistent improvements as training or model scale increases. Additionally, we explore video frames as a new source of image-text data by extracting scene-changing frames. These frames exhibit a visual distribution distinct from standard web image corpora, and models trained on this dataset achieve strong image-text retrieval performance. Overall, LAION-BVD significantly expands open access to multimodal videos at unprecedented scale and will be released to the research community for multimodal research

Modern LLMs continue to exhibit significant variance in behavior across languages, such as being able to recall factual information in some languages but not others. While typically studied as a problem to be mitigated, in this work, we propose leveraging this cross-lingual inconsistency as a tool for interpretability in mixture-of-experts (MoE) LLMs. Our knowledge localization framework contrasts routing for sets of languages where the model correctly recalls information from languages where it fails. This allows us to isolate model components that play a functional role in answering about a piece of knowledge. Our method proceeds in two stages: (1) querying the model with difficult factual questions across a diverse set of languages to generate "success" and "failure" activation buckets and then (2) applying a statistical contrastive analysis to the MoE router logits to identify experts important for knowledge. To validate the necessity of this small number of experts for answering a knowledge question, we deactivate them and re-ask the question. We find that despite only deactivating about 20 out of 6000 experts, the model no longer answers correctly in over 40% of cases. Generally, this method provides a realistic and scalable localization approach to increasingly complex LLMs, but also suggests redundant and highly dispersed knowledge parameterization in MoEs.

Large Language Models (LLMs) rely heavily on Key-Value (KV) caching to minimize inference latency. However, standard KV caches are context-dependent: reusing a cached document in a new context requires recomputing KV states to account for shifts in attention distribution. Existing solutions such as CacheBlend, EPIC, and SAM-KV mitigate this issue by selectively recomputing a subset of tokens; however, they still incur non-negligible computational overhead (FLOPs) and increased Time-to-First-Token (TTFT) latency. In this paper, we propose KV Packet, a recomputation-free cache reuse framework that treats cached documents as immutable ``packets'' wrapped in light-weight trainable soft-token adapters, which are trained via self-supervised distillation to bridge context discontinuities. Experiments on Llama-3.1 and Qwen2.5 demonstrate that the proposed KV Packet method achieves near-zero FLOPs and lower TTFT than recomputation-based baselines, while retaining F1 scores comparable to those of the full recomputation baseline. Code and experimental results are available at https://anonymous.4open.science/r/kv_packet_anonymous-46FC/.

Concept erasure is a prominent approach to achieving fairness in machine learning through the removal of sensitive attributes from learned representations. Prior concept erasure methods typically define debiasing indirectly through the failure of a chosen adversary, probe, or fairness penalty, making the target of erasure dependent on a particular decoder family or optimization setup. Rather than depending on a particular architecture, task type, or training paradigm, we introduce KINDER, which aims to directly remove sensitive-attribute information from a model's intermediate representations and is compatible with any deep model that contains an intermediate feature space. KINDER operates in a random Fourier feature space, where it estimates a sensitive subspace from protected group prototypes and projects representations onto its orthogonal complement, thereby forming an explicit representation cleaning mechanism. We conduct extensive experiments across diverse settings and show the effectiveness of KINDER on (i) unimodal and multimodal datasets, (ii) supervised and self-supervised settings, (iii) classification, regression, and image segmentation tasks, and (iv) diverse data modalities, including visual, textual, and tabular data.

Large language models (LLMs) are increasingly deployed in K–12 education, yet existing benchmarks such as C-Eval, CMMLU, GaokaoBench, and EduEval measure only whether a model can answer an exam question, i.e., factual recall. Effective educational AI further requires curriculum cognition: the structured understanding of how knowledge is organized, including prerequisite chains, concept taxonomies, experiment–concept links, and pedagogical sequencing. Curriculum cognition is neither probed by current benchmarks nor explicitly taught by current instruction-tuning data. To close this gap, we introduce K12-KGraph, a curriculum-aligned knowledge graph extracted from the official People's Education Press textbooks, covering mathematics, physics, chemistry, and biology across primary, middle, and high school, with seven node types (Concept, Skill, Experiment, Exercise, Section, Chapter, Book) and nine relation types spanning taxonomy, prerequisite, association, verification, assessment, location, and order. Building on this single graph, we derive two complementary resources: K12-Bench, a 23,640-question multi-select benchmark across five graph-derived task families (Ground, Prereq, Neighbor, Evidence, and Locate) that jointly probe curriculum cognition; and K12-Train, a KG-guided supervised fine-tuning corpus of ~2,300 QA pairs synthesized from node attributes and edge semantics. Experiments expose a clear gap and a clear remedy: on K12-Bench, even the strongest proprietary model (Gemini-3-Flash) reaches only 57% exact match and the best open-source model (Gemma-4-31B-IT) only 46%, with Prereq and Neighbor being the hardest; yet under a strictly matched 2,300-sample SFT budget on Qwen3-4B-Base and Llama3.1-8B-Base, K12-Train consistently outperforms equally sized subsets of eight mainstream instruction-tuning corpora (OpenHermes, Infinity, UltraChat, WizardLM, DataFlow, LMSYS, SmolTalk, Tulu-3) on both GaokaoBench and EduEval, showing that structural curriculum grounding is remarkably sample-efficient for educational SFT. We release the graph, benchmark, training data, and the full construction pipeline.

Large language models (LLMs) are now deployed in a wide range of real-world applications, making it important to evaluate how reliably they resist jailbreak attacks. However, existing jailbreak evaluations mainly rely on manually collected prompts or discrete text-space optimization, which limits their coverage and makes them difficult to extend to new threat settings. We present $\textbf{JailBound}$, a jailbreak evaluation framework that combines automated benchmark construction with intent-preserving attack optimization in embedding space. JailBound organizes evaluation instances with a hierarchical threat taxonomy spanning risk categories, application domains, and attack types, and uses this structure to generate meta-attack prompts with a fine-tuned meta-attack generator. It further formulates jailbreak evaluation as an embedding-space attack optimization problem and uses a first-order loss (FOL)-guided dual-branch search to jointly identify high-value vulnerable regions and safety boundary states. Under a unified evaluation protocol, we study 46 LLMs from 13 model families. The results show that JailBound covers a broader range of risk settings than existing prompt-based benchmarks, and that its optimized attacks transfer nontrivially across model families while supporting finer-grained analysis of vulnerability patterns and safety boundary behavior. $\textcolor{red}{\text{Warning: this paper includes examples that may be offensive or harmful.}}$

Articulated objects used in simulation and embodied AI are typically specified by geometry and kinematic structure, but lack the fine-grained dynamical effects that govern realistic mechanical behavior, such as frictional holding, detents, soft closing, and snap latching. Existing approaches either ignore the detailed structure of dynamics entirely, or use simple models with limited expressiveness. We introduce JODA, a framework for generating joint-level dynamics as a structured three-channel field over the joint degree of freedom, capturing conservative forces, dry friction, and damping. Instantiated using shape-constrained piecewise cubic interpolation (PCHIP), this formulation defines a compact and expressive function space that is both interpretable and compatible with differentiable simulation. Building on this representation, we develop methods for inferring and refining joint dynamics from multimodal inputs. Given visual observations and joint context, a vision-language model proposes structured dynamical primitives, which are composed into a unified dynamics field. The resulting representation supports both direct manipulation and gradient-based refinement. We demonstrate that JODA enables plausible and controllable modeling of diverse joint behaviors, providing a unified interface for inference, editing, and optimization.

Large Language Models remain vulnerable to jailbreak attacks despite extensive efforts to ensure safety alignment. Streaming generation scenarios exacerbate this vulnerability by exposing harmful tokens to users immediately upon generation, a phenomenon known as prefix exposure. Existing defenses often fail to address this real-time constraint or impose substantial latency that degrades user experience. To bridge this gap, we propose JEDI, a real-time defense method that secures streaming outputs via in-generation detection and intervention. JEDI leverages representation engineering to monitor the LLM’s internal semantic drift, using a Cumulative Sum algorithm to identify persistent, harmful intent before it manifests in the output. Upon detecting a risk, the system dynamically injects a steering vector to redirect the generation trajectory toward a safe subspace. Extensive evaluations across 6 distinct models and 11 attack vectors demonstrate that JEDI achieves defense success rates ranging from 93.6\% to 99.4\%, outperforming state-of-the-art baselines. Furthermore, JEDI preserves model utility with a Time To First Token overhead of approximately 0.003 seconds, validating its viability for latency-sensitive applications. Our code and experimental results are available at: https://anonymous.4open.science/r/JEDI-93F9

Recent one-step generative models accelerate sampling by learning deterministic flow maps of the underlying dynamics. These methods rely on learning from ordinary differential equations, leaving open how to define an exact distillation procedure for stochastic dynamics. We introduce the Itô map, an any-step stochastic flow map that takes an intermediate state and Brownian path and predicts future states in a single pass. The Itô map formulation yields novel estimators for inference-time control by providing cheap, differentiable access to posterior samples. Empirically, Itô maps produce diverse, conditionally valid endpoint samples from fixed intermediate states and support strong steering performance on synthetic and image-generation benchmarks. These results establish any-step SDE integration as a useful primitive for posterior sampling and stochastic control.

Current pre-training datasets are derived from web crawls, with all their issues, and were not designed to support mid- and post-training pipelines—for instance, they contain little explicit reasoning. Thus, many frontier labs have begun to develop their own internal datasets, starting from state-of-the-art models, to augment their pre-training data mix, e.g., with reasoning traces to address cold start problems. While demonstratively effective, none of these datasets are public, and the effect of this so-called _synthetic data_ on knowledge and skill acquisition of language models, including small ones, remains poorly understood. We present Synth, the first open-source synthetic corpus derived from 58,698 Wikipedia articles that collapses pre-, mid-, and post-training into a single training stage via structured amplification of curated encyclopedic seeds. We evaluate Synth by training a series of models: a 56M tiny model (Monad), 0.3B--0.6B dense models (Baguettotron), and a 13B-total / 1B-active Mixture-of-Experts. At iso-compute, Synth outperforms filtered web data, and our models remain competitive with similarly-sized open-weight baselines. Because Synth is back-translated from grounded passages, Synth-trained models achieve high factual precision despite 10-140$\times$ fewer training tokens, with memorization targeted by the seed corpus. These results show that synthetic datasets, including our Synth dataset, are capable of producing competitive generalist models at a significantly lower cost, enabling rapid iteration as the frontier advances. These findings open up possibilities for both generalist models with significantly increased data efficiency, as well as domain-specific models where no instruction or conversational data is available. Finally, we publicly release our Synth dataset and the series of Baguettotron models under a permissive license, thus supporting open-source language model development.

Many sequential Bayesian methods maintain a whitened operator $A_w = L^{-1}ML^{-\top}$, where $L$ is the Cholesky factor of a kernel, precision, or curvature matrix. When the underlying matrix changes by a rank-$1$ perturbation, existing techniques propagate only one-sided quantities $L^{-1}B$; refreshing the sandwiched operator $A_w$ requires a full $O(r^3)$ re-factorization that dominates iteration cost, consuming up to 91% of per-step time in natural gradient methods. We show that this cost is unnecessary. For pre-whitened inputs, intermediate factors of $L$ cancel exactly from both sides of the augmented Givens rotations, and the updated $(L^+)^{-1}M^+(L^+)^{-\top}$ can be recovered in $O(r^2)$ time by applying the rotations double-sidedly to a zero-padded matrix, followed by a rank-1 correction. We provide a self-contained derivation, stability bounds, and implementations as a C++ extension, MLX kernel, and Triton GPU kernel. On streaming sparse GP regression, online Bayesian optimization on protein fitness landscapes, and natural gradient preconditioning, the method matches full re-factorization in accuracy with up to $5.4\times$ speedups. At $r = 1024$, training time drops from 1075s to 198s, making larger curvature charts affordable and producing strictly better test MSE.

Large language models (LLMs) provide a powerful reasoning backbone for speech understanding, but integrating continuous acoustic signals into a frozen LLM remains challenging. Existing speech-to-LLM interfaces typically operate at two extremes: either enforcing near-discrete token alignment, which benefits transcription but loses paralinguistic information, or learning unconstrained continuous representations, which can drift away from the LLM’s input space and degrade autoregressive decoding. In this work, we propose Convex Gate (C-Gate), a speech–LLM bridge that constrains all speech representations to lie within the LLM’s input embedding manifold with an architectural convex-hull constraint. Concretely, each frame is represented as a convex combination of token embeddings, ensuring compatibility with the pretrained LLM while preserving continuous expressivity. Across automatic speech recognition (ASR) and emotion recognition, C-Gate achieves strong joint performance, improving LibriSpeech WER by up to 48.7\% relative while matching or exceeding single-task emotion accuracy. Beyond performance, our analysis reveals a key insight: information is not carried by discrete token identities, but by time-resolved trajectories in the embedding space. Causal interventions confirm that both the trajectory structure and alignment to the pretrained embedding manifold are critical for performance. These results suggest that geometry, rather than token discreteness, is the fundamental design factor in speech–LLM interfaces, and provide a controlled regime for studying multimodal integration in frozen LLMs. We release the checkpoint, per-sample outputs, mechanism dumps, and intervention suite for replication.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable effectiveness in general video understanding. Human motion understanding, a dominant topic in video analytics, however, still remains as a bottleneck due to the lack of explicit structural cues in 2D visual patches and the significant motion information loss caused by sparse visual temporal sampling. To address this, we propose Iris (Integrating representations of in-the-wild skeletons), a novel pose-augmented video MLLM tailored for human-centric motion understanding. Iris adopts an asymmetric dual-stream architecture, pairing the traditional sparse vision stream with a lightweight, high-frequency human pose stream to provide spatially structured and temporally dense kinematic signals. To enable pose-vision modality correspondence awareness, we design a cross-modal spatiotemporal binding mechanism, featuring a pose-structured 3D RoPE for precise temporal alignment and a person-grounded cross-attention module for explicit spatial visual grounding. Furthermore, to enable robust large-scale training, we construct an automated pose curation pipeline that organizes pose data into a delicate heterogeneous representation. Extensive experiments demonstrate that Iris achieves superior performance on multiple fine-grained motion benchmarks, while also maintaining highly efficient token computation overhead. Code will be available upon publication.

Post-hoc calibration aligns a classifier's predicted confidences with its empirical accuracy without retraining. An ideal calibrator should correct nonlinear miscalibration, scale gracefully to large label spaces, and preserve the original predictions; existing methods typically violate at least one of these properties---temperature scaling lacks expressivity, more flexible parametric alternatives introduce parameters that grow with the number of classes $C$, and other expressive methods do not preserve the rank ordering of class scores and may alter the predicted class. We propose \textbf{Invertible Logits Transformation (InvLT)}, which applies a learned scalar MLP $f:\mathbb{R}\to\mathbb{R}$ element-wise to the pre-softmax logits. Sharing $f$ across all logit dimensions makes the parameter count independent of $C$. Monotonicity of $f$---and hence preservation of the argmax prediction---is softly encouraged via a paired inverse network rather than enforced through the numerical integration required by prior monotone calibrators; this avoids their computational overhead while empirically preserving the original classification accuracy in every setting we evaluate. Across standard image classification benchmarks and a range of architectures, InvLT consistently outperforms a broad set of post-hoc baselines on standard calibration metrics.

Image-free zero-shot classifier expansion aims to extend recognition to unseen classes by establishing an association mapping between class semantics and visual classifier weights, without accessing original images. However, most existing methods rely on an association-based optimization mechanism, which struggles to meet the demands for fine-grained classification of visually and semantically similar classes. Since co-occurring attributes dominate class semantics, the model gravitates towards co-occurring attributes rather than discriminative attributes. This leads to geometrically entangled classifier weights, resulting in ambiguous decision boundaries. To address this problem, this paper proposes a novel approach that reframes weight generation through the lens of causal intervention. The method first employs comparative reasoning to disentangle discriminative and co-occurring attributes. Subsequently, it synthesizes intervened semantics to impose invariance and sensitivity constraints. This mechanism compels the model to suppress co-occurring attributes while enhancing sensitivity to discriminative attributes. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches, yielding clearer decision boundaries in fine-grained settings.