NeurIPS 2026 论文列表

Although audio-driven human animation has achieved impressive realism, it still lacks effective Multi-Scene Temporal Control: orchestrating background transitions at precise timestamps while preserving a coherent foreground subject. We show that this challenge stems from the diffusion transformer's self-attention mechanism, and addressing it requires two capabilities: (1) Semantic-Aware Temporal Decoupling, which isolates background attention across scenes while retaining cross-scene foreground attention for subject consistency; and (2) Foreground Motion Localization, which accurately tracks the foreground subject across latent frames, especially under large motion. To address these challenges, we introduce SceneShifter, a training-free framework for precise multi-scene control in audio-driven human animation. SceneShifter guides spatio-temporal self-attention by suppressing cross-scene attention among background tokens while preserving cross-scene attention among foreground tokens. To support this guidance under large motion, SceneShifter extracts dynamic foreground masks from the self-attention layers and heads that best capture subject trajectories. We further introduce SceneShifterBench, a benchmark designed to evaluate scene-timing accuracy, foreground preservation, and visual quality under large motion, multiple subjects, and occlusions. Experiments show that SceneShifter achieves frame-level scene timing, strong subject preservation, and high visual quality, outperforming existing baselines. Code and examples can be found at https://anonymous.4open.science/r/SceneShifter_neurips_26.

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.

Many public buildings provide floorplans with a “you are here” indicator to help visitors orient themselves. Floorplan localization seeks to computationally replicate this capability by determining where visual observations were captured within a floorplan. However, existing methods typically assume controlled small-scale environments and precise vectorized floorplans, limiting their ability to operate in large-scale buildings and rasterized floorplans. In this work, we present an approach for performing floorplan localization in the wild by grounding the task in a reconstructed 3D representation of the scene. Given an unconstrained image collection, our method reconstructs a gravity-aligned 3D scene and projects it into a 2D density map that serves as a floorplan proxy. Floorplan localization is then formulated as aligning this proxy with the input floorplan via a 2D similarity transform. To bridge the appearance gap between density maps and architectural floorplans, we adapt a 2D foundation model to learn cross-modal correspondences, introducing a fine-tuning scheme that encourages semantically aligned matches while preserving structural consistency. Extensive experiments demonstrate substantial improvements over prior methods, including in extremely sparse settings with as little as a single input image. Our code and data will be publicly available.

Linear mode connectivity (LMC) provides a promising foundation for understanding and merging independently trained neural networks, but existing methods typically optimize the interpolation path from only one model endpoint, limiting their scalability and effectiveness for large pretrained transformers. We propose a novel and scalable framework for enabling LMC-based model merging to {\em billion-parameter pretrained transformers}. Our method applies properly parameterized functionality-preserving weight transformations to align functionally equivalent solutions, and introduces a dual learning procedure in which both models jointly learn their corresponding transformations toward a shared linear interpolation path. This bidirectional optimization substantially reduces interpolation barriers and enables more reliable merging across large-scale architectures. Empirically, we show that our approach achieves near-zero loss barriers on WikiText for language models with medium-sized parameters, representing, to our knowledge, the first demonstration of near-barrier-free linear connectivity at this scale. In the vision domain, ViT-L maintains above 69\% ImageNet top-1 accuracy throughout the interpolation path, while modern billion-parameter LLMs exhibit only small loss barriers. These results suggest that properly resolving parameter symmetries enables large pretrained Transformers to be connected and merged through simple linear paths with substantially improved interpolation performance.

Despite surpassing human performance across mathematics, coding, and other knowledge intensive tasks, large language models (LLMs) continue to struggle to causally reason. A core obstacle is the target data itself: causal systems are complex and often expressed in non-executable forms, and ground-truth answers to causal queries are inherently scarce. We introduce, **CauSim**, a framework that turns causal reasoning from a scarce-label problem into a scalable, supervised one. It constructs \textit{increasingly complex causal simulators}: executable structural causal models (SCMs), incrementally built by LLMs, that scale to globally complex systems while maintaining verifiable answers to any causal query. **CauSim** operates across representations by formalizing non-executable causal knowledge into code, allowing for data augmentation, and informalizing executable SCMs into natural language, enabling supervision in previously unsupervisable representations. We structure our research into two parts: (1) how to construct increasingly complex causal simulators, and (2) a systematic study of what **CauSim** enables, demonstrating generalization across representations, consistent gains from curriculum scaling and data volume, LLM self-improvement through self-generated simulators, and data augmentation via formalization of existing domain knowledge.

Post-training weight quantization is crucial for reducing the deployment cost of large language models (LLMs), yet maintaining model quality in the ultra-low-bit regime ($\le$ 3 bits) remains challenging due to highly non-uniform weight sensitivity and the lack of principled precision allocation strategies. In this work, we show that the weight sensitivity is inherently dynamic and depends on the current quantized model, rather than the full-precision reference used in existing methods. Leveraging this observation, we introduce a sensitivity estimate that better captures the changes in marginal loss during progressive quantization. We further uncover a structured pattern in sensitivity distribution, revealing a bi-directional concentration across input and output channels, which enables hardware-aligned yet expressive partitioning via channel reordering. Building on these insights, we propose ScaleBITS, a scalable framework that combines structured partitioning and an efficient approximation to greedy allocation. Our approach enables fine-grained, global bitwidth search while preserving hardware efficiency. Experiments show that ScaleBITS significantly improves over uniform-precision quantization (up to +36%) and outperforms state-of-the-art sensitivity-aware baselines (up to +13%) in the ultra-low-bit regime, without adding runtime overhead.

A post-hoc Bayesian procedure should not change its evidence or uncertainty estimates under a function-preserving reparameterization. Standard empirical-Bayes (EB) Laplace approximation violates this principle for ReLU networks. Positive homogeneity induces continuous scale orbits of equivalent parameter vectors, yet the standard isotropic objective depends on the chosen representative through the prior norm and curvature log-determinant term. Equivalent ReLU parameterizations can thus yield different marginal likelihoods, selected prior precisions, posterior covariances, and linearized predictive uncertainties. We fix this by introducing $\gamma$-canonicalization, selecting a canonical representative of each ReLU scale orbit and fitting Laplace there. This is equivalent to transporting the isotropic Gaussian prior from the canonical representative back to the original coordinates, producing a precision field that transforms compatibly with the curvature. For every $\gamma\in[0,1]$, the generalized evidence and linearized predictive distribution are invariant under ReLU rescaling. The parameter $\gamma$ indexes a family of orbit-consistent priors and is selected jointly with the prior precision by marginal likelihood, making EB Laplace a well-defined procedure on ReLU scale-equivalence classes. Empirically, we show that standard EB-Laplace varies substantially across functionally identical ReLU rescalings, while $\gamma$-canonicalized Laplace collapses this variation to zero.

Image editing aims to modify specific attributes of an image while preserving all other aspects. In practice, however, applying even simple editing instructions to existing methods often leads to unintended changes. We identify the central issue: a fundamental causal principle of \emph{minimal change}, which requires that an intervention alter only the intended attributes in the output while leaving all others invariant, has not been explicitly incorporated as an optimization objective. To encourage minimal change, prior work rooted in causal representation learning typically imposes an $L_1$ regularizer on the latent difference between pre- and post-edit representations. However, this strategy does not scale to modern image editing models: $L_1$ regularization fails to induce true sparsity and instead tends to shrink all latent differences uniformly; sparsity in the latent space does not translate to localized change in the output of nonlinear deep image generation models; and the multi-domain or counterfactual supervision it requires is generally unavailable. To effectively leverage the minimal change principle for controllable image editing, we cast it as a reinforcement learning objective. Rather than constraining latent representations, we design rewards that directly capture minimal change over intervention outcomes and optimize them explicitly during training. To produce these rewards reliably, we further introduce an agentic reward model that leverages the chain-of-thought reasoning ability of vision-language models without requiring human annotations. The idea is that instead of emitting a single, unreliable scalar score, the model reasons explicitly about two complementary failure modes, unimplemented changes and unintended changes, and produces this supervision automatically. Experimental results demonstrate that our method produces significantly more localized and semantically consistent edits compared to existing approaches, reducing unintended changes.

Vision-language-action (VLA) policies increasingly generate natural-language rationales before executing embodied actions. These rationales are attractive as monitors for safety-critical behavior because users or automated checkers can inspect whether the model appears to understand the scene and intend a safe action before execution. We show that this signal can be misleading. In many reasoning VLAs, the rationale generator and action head share continuous hidden states, while the generated text reveals only part of that representation. This creates a Say the Same, Act Differently failure mode, where rationale text remains stable while the action output changes substantially. We first study this phenomenon by introducing a new statistical method to separate text-sensitive and action-sensitive directions in hidden-state space of a reasoning VLA. Across the driving VLA Alpamayo and the manipulation VLA InstructVLA, we show that the dominant action direction lies largely outside the active text span, with mean projection ratios of only 1.9% and 5.3%. We define the residual as a text-neutral action-subspace (TNAS) direction, which tests whether action-relevant directions remain after removing measured text-sensitive directions. We then use TNAS to guide white-box pixel-space projected-gradient-descent (PGD) attacks on Alpamayo. TNAS-guided PGD successfully uses a localized two-forward-camera patch to induce a 30.5 m mean ADE shift in trajectory planning while preserving the exact rationale in 86.7% of cases. Moreover, across multiple text-preserving PGD objectives, the optimization direction consistently aligns with TNAS. These results reveal a fundamental shared-cache monitorability gap in reasoning VLAs: stable rationale text should not be treated as sufficient evidence that the action output remains stable.

Urban uncrewed aerial vehicle (UAV) vision-language navigation (VLN) requires agents to follow instructions across extended urban spaces, inherently demanding long-term memory and geospatial grounding. However, scaling existing benchmarks remains difficult because of their reliance on costly reconstructed 3D assets, limiting geographic diversity and episode scale. To address this, we introduce SatNav, a scalable, long-horizon UAV VLN benchmark built from high-resolution satellite imagery. SatNav targets city-level navigation missions and uses satellite crops as approximations of UAV nadir views for visual observations. Through an automated cue-to-episode pipeline, SatNav constructs 118K episodes from 59 scenes across 18 cities, with an average trajectory length of 379 m. To stress-test long-horizon memory and geospatial reasoning, SatNav defines three task families: Boundary, Landmark, and Route, targeting loop progress tracking, landmark-based spatial grounding, and route following with counting cues. Benchmarking classical and recent large vision language model (LVLM) based VLN agents on SatNav shows that city-scale navigation remains challenging. We further introduce SwiftVLN, a modular framework with switchable memory components, and conduct systematic memory-design ablations. Finally, satellite-to-UAV transfer experiments show that satellite-trained navigation models can operate on real-flight UAV observations, showing the practical relevance of SatNav.

While foundation models have significantly advanced human recognition across diverse modalities, they predominantly rely on static, geometric feature extraction. This approach fundamentally diverges from human perception. Consequently, current models often suffer from ``semantic blindness,'' overfitting to transient noise while failing to leverage invariant soft biometrics, and struggle to capture temporal motion signatures. To bridge this gap, we propose SapiensID 2.0, a human recognition framework enriched with both semantic and temporal awareness. To overcome the lack of soft-biometric annotations, we transfer zero-shot semantic knowledge from Multimodal Large Language Models (MLLMs) into a discriminative embedding space. We resolve the dimensional mismatch between these spaces using Invariant Trait Alignment (ITA) to distill core persistent traits, and Transient Noise Disentanglement (TND) to decouple artifacts like clothing. Furthermore, we design a Kinematic Semantic Attention Head (K-SAH) that extends spatial attention across temporal windows. By tracking semantic patches over time, K-SAH captures rich kinematic signatures without requiring large-scale video datasets. Extensive experiments demonstrate that SapiensID 2.0 achieves state-of-the-art performance across image- and video-based person re-identification and gait recognition, while maintaining robust face recognition capabilities.

As multimodal LLMs increasingly emphasize video and audio, a common assumption is that solving such tasks calls for native omnimodal models. We show that this is not always necessary: coding agents equipped with only text+vision and a sandboxed tool-using interface can perform competitively with, and in several settings outperform, SOTA native omnimodal models and predefined multimodal agent scaffolds across multiple audio-video benchmarks. Our trajectory analysis suggests that their advantage comes from coding agents writing code and orchestrating tools to retrieve relevant content from transcripts, frames, and other non-textual modality signals from raw inputs. This effectively converts omnimodal tasks into evidence retrieval and information processing problems, avoiding the inefficiency of ingesting entire videos or audio streams into context. To characterize remaining limitations, we propose a failure taxonomy and a process-level analysis of tool-use traces, and find that simple skill injection, including human-written skills and self-distilled skills from execution logs, can substantially improve performance over the no-skill baseline. To examine whether such capability can be elicited in open-source models, we further introduce Code-X, a complete training recipe with the OmniCoding trajectory dataset and verifiable reward, providing an exploratory baseline on Qwen-3.5-9B and Qwen-3.6-27B. Finally, given the maturity of many-modality understanding, we argue that the more meaningful frontier lies in many-modality processing, and introduce TerminalBench-O, the first process-level benchmark designed for coding agents on real-world omnimodal processing tasks. Together, our findings open new directions for omnimodal content processing and evaluation.

Vision-language models map images and text into a joint embedding space. However, these embeddings often entangle multiple semantic features, which limits their interpretability and controllability. While sparse autoencoders have emerged as a useful tool for decomposing these embeddings into monosemantic features, their application to joint embedding spaces has largely relied on an implicit, untested assumption that semantically corresponding features share the same directions across modalities. In this paper, we challenge this assumption by identifying discrepancies in feature directions for the same concept across image and text modalities, a phenomenon we term *cross-modal feature heterogeneity*. We demonstrate that this heterogeneity is a key driver of the *modality split*, where a shared concept activates different latents depending on the modality. This finding further reveals why aligning latent activations alone is insufficient to resolve the underlying feature mismatch. To address this misalignment, we propose an approach that trains sparse autoencoders to preserve the unique feature geometry of each modality and aligns corresponding features post hoc. Our method improves reconstruction fidelity and enhances performance in cross-modal retrieval and concept steering.

A vision-language (VL) agent wraps a frozen vision-language model (VLM) into a test-time system that may call external tools during execution before producing a response. The central problem is how the VLM and its tools interact across multiple steps. Two recipes are common, both rigid. Thinking for longer (borrowed from text-only agents) sharpens reasoning, but without in-loop visual refresh, it lets visual evidence decay into hallucination. Querying tools more often, as traditional agents do, treats outputs as ground truth, a rigid commitment that fails when calls are noisy. We propose Saliency-Aware Principle Selection (SAP), a training-free inference-time method that avoids both rigidities: it searches a broad space of high-level principles, short textual directives (e.g., ``re-examine the image whenever an intermediate conclusion is formed'') that each prescribe a different way for the VLM to weigh tool advice across steps, and selects the principle whose parallel reasoning routes best fit the actual visual evidence. Tool outputs are treated as advisory (consulted, never authoritative), and a population-based evolutionary loop drives the search. Under the same token budget as long chain-of-thought, SAP substantially reduces object hallucination while remaining competitive on reasoning-heavy benchmarks. Being training-free and composed of independent routes, SAP is plug-and-play on any VLM and parallel for modern agent deployment.

Multimodal large language models (MLLMs) often fail to transfer safety capabilities learned in the text modality to semantically equivalent non-text inputs, revealing a persistent multimodal safety gap. We study this gap from a representation-geometric perspective by analyzing a text-aligned refusal direction and a modality-induced drift direction. We show that multimodal inputs compress the usable separation along the refusal direction, making it no longer reliable for identifying and refusing harmful inputs. We refer to this failure mode as Safety Geometry Collapse. We quantify it through conditional refusal separability and show that stronger modality-induced drift is consistently associated with weaker refusal separability and higher attack success rates. We then validate the causal role of modality-induced drift through a fixed-strength activation intervention: counteracting the estimated drift restores refusal separability and improves multimodal safety. After drift correction, we further observe self-rectification, where the model recovers its ability to recognize and refuse harmful multimodal inputs during forward dynamics. This effect also provides an internal signal of the model’s perceived harmfulness of each input. Motivated by this signal, we propose ReGap, a training-free inference-time method that adaptively corrects modality drift using self-rectification. Experiments across multiple multimodal safety benchmarks and utility benchmarks demonstrate the effectiveness of ReGap, which significantly improves the safety of MLLMs without compromising general capabilities. Our findings highlight representation-level modality alignment as a crucial direction for real-time safety improvement and for building safer, more reliable MLLMs.

Source-Free Domain Adaptation (SFDA) aims to adapt a source-pretrained model to an unlabeled target domain without access to the original source domain. While early single-model approaches rely on self-refinement, they are inherently susceptible to confirmation bias and struggle to correct their own systematic errors. To overcome this limitation, recent methods introduce Vision-Language (ViL) models as external knowledge sources. However, these approaches operate in a largely unidirectional paradigm, using the ViL model primarily to supervise the source-pretrained model. This overlooks a key structural property: the two models exhibit distinct failure modes -- where one produces an incorrect prediction, the other may produce a correct one, creating a natural opportunity for mutual correction within the target domain. Yet, without ground-truth labels, identifying which model is correct on any given sample is non-trivial, and naively exchanging predictions risks propagating errors across models. To address this challenge, we propose SafeCut, a novel approach that leverages the cut statistic as a label-free measure of prediction reliability to gate cross-model supervision. Our approach dynamically controls both the direction and strength of supervision based on relative reliability, selectively amplifying true corrections while suppressing miscorrections on a per-sample basis. We further provide theoretical justification showing that this reliability-gated mechanism guarantees a net-positive correction signal. Extensive experiments across diverse SFDA benchmarks demonstrate that SafeCut achieves state-of-the-art performance, highlighting the effectiveness of safeguarding mutual correction in SFDA via cut statistics. Codes are available https://anonymous.4open.science/r/SafeCut_Source_Free_Domain_Adaptation-65D0/

Large language models (LLMs) have shown strong performance on biomedical tasks, yet evaluating their reasoning in safety-critical pharmacological contexts requires large-scale, structured datasets. We present SafeDrug, a benchmark dataset for systematic evaluation of pharmacological reasoning across multiple task dimensions. SafeDrug integrates heterogeneous sources, including adverse drug event records, drug–drug interaction knowledge, literature-derived evidence, and population-specific information such as child, adult, and older adult groups, into a unified reasoning-oriented QA format. The benchmark comprises two complementary subsets. \textbf{SafeDrug-large} provides broad coverage at scale, while \textbf{SafeDrug-small} is a high-quality human-annotated subset for precise evaluation. It supports tasks such as prediction, risk assessment, mechanism explanation, evidence-grounded QA, and drug replacement across both single-drug and multi-drug settings. Our evaluation shows that LLMs capture outcome-level associations but struggle with mechanistic reasoning, evidence alignment, and population-aware predictions. SafeDrug provides a foundation for reproducible benchmarking and advances research on evidence-grounded and population-aware pharmacological reasoning in LLMs.

Multimodal large reasoning models can reason over images and text, but they remain vulnerable to multimodal jailbreaks. Recent safety-alignment methods reduce attack success rate (ASR) by fine-tuning on reasoning traces from stronger teacher models. We show that this has a hidden cost of over-refusal. The aligned model learns to reject benign inputs that contain sensitive words or visual cues. This is especially harmful for boundary-safe inputs, where the prompt looks risky on the surface but is safe in context. We identify external teacher supervision as a key factor in this behavior. It introduces a distributional mismatch and shifts supervision away from the base model's own reasoning distribution. To address this, we propose SAGE (Safety-Aware Guided Elicitation), a self-guided data generation framework for multimodal safety alignment. SAGE guides generation in two stages: a reasoning-level guide first elicits a safety-aware trace, and a decision-level guide then elicits the final response conditioned on that trace. This factorized guidance lets SAGE shape both the reasoning process and the final safety decision without relying on external teacher traces. On LLaVA-CoT, SAGE reduces FigStep ASR from 86.2% to 2.1% and achieves an average over-refusal of 43.1, compared with 66.9--71.1 for prior reasoning-based safety-alignment methods, while preserving general utility. We will release the SAGE training dataset.

Software performance optimization is a notoriously complex and manual task. Despite the growing use of Large Language Models (LLMs) for code refinement, we still lack benchmarks that capture how optimization actually happens in real-world codebases. Existing frameworks often oversimplify the problem by focusing on isolated functions or a single performance metric, missing the critical trade-offs between execution time and memory footprint, the inherent noise of the measurement environment, and the variability introduced by different input data and execution conditions. We address this by introducing SWE-Pro, a repository-level benchmark derived from 102 expert-written optimizations from open-source projects. Unlike previous benchmarks, SWE-Pro pairs each task with parameterized tests to evaluate runtime, peak memory, and Time-Weighted Memory Usage (TWMU) varying input data and execution conditions under noise-aware measurement conditions. Our evaluation shows that current LLMs struggle significantly: runtime gains are negligible, and memory optimizations are nearly non-existent. This stands in sharp contrast to expert implementations, which achieve an aggregate speedup of 15.5x and peak memory reduction of 171.3x over benchmark tasks. Expert-written improvements are observed in 91.2% of tasks for runtime and 65.7% for peak memory. Our findings expose a substantial gap between current LLMs capabilities and the demands of expert-level engineering.