NeurIPS 2025 论文列表

Speculative decoding improves LLM inference by generating and verifying multiple tokens in parallel, but existing systems suffer from suboptimal performance due to a mismatch between dynamic speculation and static runtime assumptions. We present Yggdrasil, a co-designed system that enables latency-optimal speculative decoding through context-aware tree drafting and compiler-friendly execution. Yggdrasil introduces an equal-growth tree structure for static graph compatibility, a latency-aware optimization objective for draft selection, and stage-based scheduling to reduce overhead. Yggdrasil supports unmodified LLMs and achieves up to $3.98\times$ speedup over state-of-the-art baselines across multiple hardware setups.

Commuting Origin-destination (OD) flows, capturing daily population mobility of citizens, are vital for sustainable development across cities around the world. However, it is challenging to obtain the data due to the high cost of travel surveys and privacy concerns. Surprisingly, we find that satellite imagery, publicly available across the globe, contains rich urban semantic signals to support high-quality OD flow generation, with over 98\% expressiveness of traditional multisource hard-to-collect urban sociodemographic, economics, land use, and point of interest data. This inspires us to design a novel data generator, GlODGen (Global-scale OriginDestination Flow Generator), which can generate OD flow data for any cities of interest around the world. Specifically, GlODGen first leverages Vision-Language Geo-Foundation Models to extract urban semantic signals related to human mobility from satellite imagery. These features are then combined with population data to form region-level representations, which are used to generate OD flows via graph diffusion models. Extensive experiments on 4 continents and 6 representative cities show that GlODGen has great generalizability across diverse urban environments on different continents and can generate OD flow data for global cities highly consistent with real-world mobility data. We implement GlODGen as an automated tool, seamlessly integrating data acquisition and curation, urban semantic feature extraction, and OD flow generation together. It has been released at https://github.com/tsinghua-fib-lab/generate-od-pubtools.

Large Vision-Language Models (VLMs) have achieved remarkable success in understanding complex real-world scenarios and supporting data-driven decision-making processes. However, VLMs exhibit significant vulnerability against adversarial examples, either text or image, which can lead to various adversarial outcomes, e.g., jailbreaking, hijacking, and hallucination, etc. In this work, we empirically and theoretically demonstrate that VLMs are particularly susceptible to image-based adversarial examples, where imperceptible perturbations can precisely manipulate each output token. To this end, we propose a novel attack called Vision-language model Manipulation Attack (VMA), which integrates first-order and second-order momentum optimization techniques with a differentiable transformation mechanism to effectively optimize the adversarial perturbation. Notably, VMA can be a double-edged sword: it can be leveraged to implement various attacks, such as jailbreaking, hijacking, privacy breaches, Denial-of-Service, and the generation of sponge examples, etc, while simultaneously enabling the injection of watermarks for copyright protection. Extensive empirical evaluations substantiate the efficacy and generalizability of VMA across diverse scenarios and datasets. Code is available at https://github.com/Trustworthy-AI-Group/VMA.

Hallucination remains a major challenge in multimodal large language models (MLLMs). To address this, various contrastive decoding (CD) methods have been proposed that contrasts original logits with hallucinated logits generated from perturbed inputs. While CD has shown promise in vision-language models (VLMs), it is not well-suited for AV-LLMs, where hallucinations often emerge from both unimodal and cross-modal combinations involving audio, video, and language. These intricate interactions call for a more adaptive and modality-aware decoding strategy. In this paper, we propose Audio-Visual Contrastive Decoding (AVCD)—a novel, training-free decoding framework designed to model trimodal interactions and suppress modality-induced hallucinations in AV-LLMs. Unlike previous CD methods in VLMs that corrupt a fixed modality, AVCD leverages attention distributions to dynamically identify less dominant modalities and applies attentive masking to generate perturbed output logits. To support CD in a trimodal setting, we also reformulate the original CD framework to jointly handle audio, visual, and textual inputs. Finally, to improve efficiency, we introduce entropy-guided adaptive decoding, which selectively skips unnecessary decoding steps based on the model’s confidence in its predictions. Extensive experiments demonstrate that AVCD consistently outperforms existing decoding methods. Especially, on the AVHBench dataset, it improves accuracy by 2% for VideoLLaMA2 and 7% for video-SALMONN, demonstrating strong robustness and generalizability. Our code is available at : https://github.com/kaistmm/AVCD.

Edge deployment of large Vision-Language Models (VLMs) increasingly relies on flash-based weight offloading, where activation sparsification is used to reduce I/O overhead. However, conventional sparsification remains model-centric, selecting neurons solely by activation magnitude and neglecting how access patterns influence flash performance. We present Neuron Chunking, an I/O-efficient sparsification strategy that operates on *chunks*—groups of contiguous neurons in memory—and couples neuron importance with storage access cost. The method models I/O latency through a lightweight abstraction of access contiguity and selects chunks with high utility, defined as neuron importance normalized by estimated latency. By aligning sparsification decisions with the underlying storage behavior, Neuron Chunking improves I/O efficiency by up to 4.65× and 5.76× on Jetson Orin Nano and Jetson AGX Orin, respectively.

Vision-based intrusion detection has multiple applications in practical scenarios, e.g., autonomous driving, intelligent monitoring, and security. Previous works mainly focus on improving the intrusion detection performance, without a comprehensive and in-depth understanding of the intrusion scene. To fill this gap, we explore a novel task called Multimodal Large Language Models based Intrusion Scene Understanding (MLLM-ISU) and report a comprehensive benchmark for the task. Specifically, we first design an effective and automatic visual question-answer generation strategy, constructing a new MLLM-ISU dataset, with 3000 VQA evaluation Pairs, 8925 training Pairs, and six relevant subtasks. Then, we perform a comprehensive assessment on various state-of-the-art proprietary and open-source MLLMs, e.g., DeepSeek-VL2, GPT-4o, Qwen2.5-VL, etc, and find that current MLLMs have weak abilities for this task. Further, in order to improve the intrusion understanding capabilities of current MLLMs, we propose a Post-Training Framework with three sequential training stages, i.e., Intrusion-aware Visual Instruction Pre-training, Intrusion Chain of Thought tuning, and Intrusion-centric VQA tuning, and sufficient experiments and comparisons are conducted to verify the effectiveness of the proposed three-stages training framework. Available datasets and codes: https://github.com/1012537710/MLLM-ISU.

Emotion recognition based on physiological signals plays a vital role in psychological health and human–computer interaction, particularly with the substantial advances in multimodal emotion recognition techniques. However, two key challenges remain unresolved: 1) how to effectively model the intra-modal long-range dependencies and inter-modal correlations in multimodal physiological emotion signals, and 2) how to address the performance limitations resulting from missing multimodal data. In this paper, we propose a multimodal bidirectional Mamba (BiMamba) network with test-time adaptation (TTA) for emotion recognition named BiM-TTA. Specifically, BiM-TTA consists of a multimodal BiMamba network and a multimodal TTA. The former includes intra-modal and inter-modal BiMamba modules, which model long-range dependencies along the time dimension and capture cross-modal correlations along the channel dimension, respectively. The latter (TTA) mitigates the amplified distribution shifts caused by missing multimodal data through two-level entropy-based sample filtering and mutual information sharing across modalities. By addressing these challenges, BiM-TTA achieves state-of-the-art results on two multimodal emotion datasets.

Multi-modal models are data hungry. While datasets with natural images are abundant, medical image datasets can not afford the same luxury. To enable representation learning for medical images at scale, we turn to YouTube, a platform with a large reservoir of open-source medical pedagogical videos. We curate MedicalNarratives, a dataset 4.7M medical image-text pairs, with 1M samples containing dense annotations in the form of traces spatial traces (and bounding boxes), and 118K videos centered on the trace event (with aligned text), enabling spatiotemporal grounding beyond single frames. Similar to think-aloud studies where instructors speak while hovering their mouse cursor movements over relevant image regions, 1M images in MedicalNarratives contains localized mouse traces in image pixels, creating a spatial association between the text and pixels. To evaluate the utility of MedicalNarratives, we train GenMedClip with a CLIP-like objective using our dataset spanning 12 medical domains. GenMedClip outperforms previous state-of-the-art models on all 12 domains on a newly constructed medical imaging benchmark. Data, demo, code, and models will be made available.

Reinforcement fine-tuning (RFT) has shown great promise in achieving humanlevel reasoning capabilities of Large Language Models (LLMs), and has recently been extended to MLLMs. Nevertheless, reasoning about videos, which is a fundamental aspect of human intelligence, remains a persistent challenge due to the complex logic, temporal and causal structures inherent in video data. To fill this gap, we propose VideoRFT, a novel approach that extends the RFT paradigm to cultivate human-like video reasoning capabilities in MLLMs. VideoRFT follows the standard two-stage scheme in RFT: supervised fine-tuning (SFT) with chain-of-thought (CoT) annotations, followed by reinforcement learning (RL) to improve generalization. A central challenge to achieve this in the video domain lies in the scarcity of large-scale, high-quality video CoT datasets. We address this by building a multi-expert-driven, cognition-inspired CoT curation pipeline. First, we devise a cognition-inspired prompting strategy to elicit a reasoning LLM to generate preliminary CoTs based solely on rich, structured, and literal representations of video content. Subsequently, these CoTs are revised by a MLLM conditioned on the actual video, ensuring visual consistency and reducing visual hallucinations. This pipeline results in two new datasets, i.e.VideoRFT-CoT-102K for SFT and VideoRFT-RL-310K for RL. To further strengthen the RL phase, we introduce a novel semantic-consistency reward that explicitly promotes the alignment between textual reasoning and visual evidence. This reward encourages the model to produce coherent, context-aware reasoning outputs grounded in visual input. Extensive experiments show that VideoRFT achieves state-of-the-art performance on six video reasoning benchmarks.

Accuracy remains a standard metric for evaluating AI systems, but it offers limited insight into how models arrive at their solutions. In this work, we introduce a benchmark based on brainteasers written in long narrative form to probe more deeply into the types of reasoning strategies that models use. Brainteasers are well-suited for this goal because they can be solved with multiple approaches, such as a few-step solution that uses a creative insight or a longer solution that uses more brute force. We investigate large language models (LLMs) across multiple layers of reasoning, focusing not only on correctness but also on the quality and creativity of their solutions. We investigate many aspects of the reasoning process: (1) semantic parsing of the brainteasers into precise mathematical competition style formats; (2) self-correcting solutions based on gold solutions; (3) producing step-by-step sketches of solutions; and (4) making use of hints. We find that LLMs are in many cases able to find creative, insightful solutions to brainteasers, suggesting that they capture some of the capacities needed to solve novel problems in creative ways. Nonetheless, there also remain situations where they rely on brute force despite the availability of more efficient, creative solutions, highlighting a potential direction for improvement in the reasoning abilities of LLMs.

Large language models (LLMs) have shown promise in transforming machine learning research, yet their capability to faithfully implement genuinely novel ideas from recent research papers—ideas unseen during pretraining—remains unclear. We introduce ResearchCodeBench, a benchmark that evaluates LLMs’ ability to translate cutting-edge ML contributions from top 2024-2025 research papers into executable code. We assessed 30+ proprietary and open-source LLMs, finding that even the best models correctly implement less than 40% of the code. We present empirical findings on performance comparison, contamination, and error patterns. By providing a rigorous evaluation platform, ResearchCodeBench enables continuous understanding and advancement of LLM-driven innovation in research code generation.

Pre-training vision-language representations on human action videos has emerged as a promising approach to reduce reliance on large-scale expert demonstrations for training embodied agents. However, prior methods often employ time con- trastive learning based on goal-reaching heuristics, progressively aligning language instructions from the initial to the final frame. This overemphasis on future frames can result in erroneous vision-language associations, as actions may terminate early or include irrelevant moments in the end. To address this issue, we propose Action Temporal Coherence Learning (AcTOL) to learn ordered and continuous vision-language representations without rigid goal-based constraint. AcTOL treats a video as a continuous trajectory where it (1) contrasts semantic differences be- tween frames to reflect their natural ordering, and (2) imposes a local Brownian bridge constraint to ensure smooth transitions across intermediate frames. Exten- sive imitation learning experiments on both simulated and real robots show that the pretrained features significantly enhance downstream manipulation tasks with high robustness to different linguistic styles of instructions, offering a viable pathway toward generalized embodied agents. Our project page is at https://actol-pretrain.github.io/.

Vision-Language Models (VLMs) excel at visual understanding but often suffer from visual hallucinations, where they generate descriptions of nonexistent objects, actions, or concepts, posing significant risks in safety-critical applications. Existing hallucination mitigation methods typically follow one of two paradigms: generation adjustment, which modifies decoding behavior to align text with visual inputs, and post-hoc verification, where external models assess and correct outputs. While effective, generation adjustment methods often rely on heuristics and lack correction mechanisms, while post-hoc verification is complicated, typically requiring multiple models and tending to reject outputs rather than refine them. In this work, we introduce REVERSE, a unified framework that integrates hallucination-aware training with on-the-fly self-verification. By leveraging a new hallucination-verification dataset containing over 1.3M semi-synthetic samples, along with a novel inference-time retrospective resampling technique, our approach enables VLMs to both detect hallucinations during generation and dynamically revise those hallucinations. Our evaluations show that REVERSE achieves state-of-the-art hallucination reduction, outperforming the best existing methods by up to 12% on CHAIR-MSCOCO and 34% on HaloQuest.

Generation of images containing multiple humans, performing complex actions, while preserving their facial identities, is a significant challenge. A major factor contributing to this is the lack of a a dedicated benchmark. To address this, we introduce MultiHuman-Testbench, a novel benchmark for rigorously evaluating generative models for multi-human generation. The benchmark comprises 1800 samples, including carefully curated text prompts, describing a range of simple to complex human actions. These prompts are matched with a total of 5,550 unique human face images, sampled uniformly to ensure diversity across age, ethnic background, and gender. Alongside captions, we provide human-selected pose conditioning images which accurately match the prompt. We propose a multi-faceted evaluation suite employing four key metrics to quantify face count, ID similarity, prompt alignment, and action detection. We conduct a thorough evaluation of a diverse set of models, including zero-shot approaches and training-based methods, with and without regional priors. We also propose novel techniques to incorporate image and region isolation using human segmentation and Hungarian matching, significantly improving ID similarity. Our proposed benchmark and key findings provide valuable insights and a standardized tool for advancing research in multi-human image generation.

The Transformer architecture serves as the foundation of modern AI systems, powering recent advances in Large Language Models (LLMs) and Large Multimodal Models (LMMs). Central to these models, attention mechanisms capture contextual dependencies via token interactions. Beyond inference, attention has been widely adopted for interpretability, offering insights into model behavior. Among interpretability techniques, attention flow --- which traces global information transfer across layers --- provides a more comprehensive perspective than single-layer attention maps. However, computing attention flow is computationally intensive due to the high complexity of max-flow algorithms. To address this challenge, we introduce FlowPrune, a novel framework that accelerates attention flow analysis by pruning the attention graph before applying max-flow computations. FlowPrune uses the Max-Flow Min-Cut Theorem and two structural properties of Transformer to identify and eliminate non-critical graph regions. It comprises two components: Edge Pruning, which removes insignificant attention edges, and Layer Compression, which discards layers with minimal contributions to the flow. We conduct extensive experiments on LLaMA and LLaVA to evaluate the robustness and effectiveness of FlowPrune. Our results show that FlowPrune achieves high agreement with the original attention flow in both absolute and relative error metrics, as well as in identifying influential input tokens. Finally, case studies in both NLP and vision domains demonstrate that FlowPrune produces consistent interpretability outcomes as the original Attention Flow, validating its practical utility. The code for this paper is publicly available.

One way to mitigate risks in vision-language models (VLMs) is to censor dangerous samples from their training data. However, data moderation can be easily bypassed when harmful images are split into small, benign-looking patches, scattered across many training samples. VLMs may then learn to piece these fragments together and generate harmful responses at inference, either from full images or text references. For instance, if trained on image patches from a bloody scene paired with the descriptions "safe," VLMs may later describe, the full image or a text reference to the scene, as "safe." We define the core ability of VLMs enabling this attack as $\textit{visual stitching}$—the ability to integrate visual information spread across multiple training samples that share the same textual descriptions. In our work, we first demonstrate visual stitching abilities in common open-source VLMs on three datasets where each image is labeled with a unique synthetic ID. We split each $(\texttt{image}, \texttt{ID})$ pair into $\{(\texttt{patch}, \texttt{ID})\}$ pairs at different granularities for finetuning, and we find that models can verbalize the correct IDs from full images or text reference. Building on this, we simulate the adversarial data poisoning scenario mentioned above by using patches from dangerous images and replacing IDs with text descriptions like "safe" or "unsafe", demonstrating how harmful content can evade moderation in patches and later be reconstructed through visual stitching, posing serious VLM safety risks.

Multimodal representation learning, exemplified by multimodal contrastive learning (MMCL) using image-text pairs, aims to learn powerful representations by aligning cues across modalities. This approach relies on the core assumption that the exemplar image-text pairs constitute two representations of an identical concept. However, recent research has revealed that real-world datasets often exhibit cross-modal misalignment. There are two distinct viewpoints on how to address this issue: one suggests mitigating the misalignment, and the other leveraging it. We seek here to reconcile these seemingly opposing perspectives, and to provide a practical guide for practitioners. Using latent variable models we thus formalize cross-modal misalignment by introducing two specific mechanisms: Selection bias, where some semantic variables are absent in the text, and perturbation bias, where semantic variables are altered—both leading to misalignment in data pairs. Our theoretical analysis demonstrates that, under mild assumptions, the representations learned by MMCL capture exactly the information related to the subset of the semantic variables invariant to selection and perturbation biases. This provides a unified perspective for understanding misalignment. Based on this, we further offer actionable insights into how misalignment should inform the design of real-world ML systems. We validate our theoretical findings via extensive empirical studies on both synthetic data and real image-text datasets, shedding light on the nuanced impact of cross-modal misalignment on multimodal representation learning.

Fine-tuning large language models (LLMs) to aggregate multiple preferences has attracted considerable research attention. With aggregation algorithms advancing, a potential economic scenario arises where fine-tuning services are provided to agents with different preferences. In this context, agents may benefit from strategically misreporting their preferences, but this could harm the aggregation performance. This paper addresses such incentive issues by framing it as a mechanism design problem: an LLM provider determines the fine-tuning objective (training rule) and the pricing scheme (payment rule) for agents. We primarily focus on training rules that maximize social welfare subject to certain regularizations, referred to as SW-Max rules. First, we show that under most circumstances, truthful reporting is sub-optimal with simply a SW-Max rule, thereby highlighting the necessity of payments. Second, we extend the VCG payment to implement SW-Max rules in dominant-strategy incentive compatibility (DSIC). We characterize sufficient conditions for payment equivalence and derive the necessary conditions for a payment rule to implement a SW-Max rule in DSIC and other principles. Third, we demonstrate that our mechanism is approximately DSIC with perturbed input, showcasing its robustness against the inevitable errors in real-world applications. Experiments on real LLM training results further confirm the practical implications of our results.

This paper presents a novel approach to addressing the long-sequence problem in high-resolution medical images for Vision Transformers (ViTs). Using smaller patches as tokens can enhance ViT performance, but quadratically increases computation and memory requirements. Therefore, the common practice for applying ViTs to high-resolution images is either to: (a) employ complex sub-quadratic attention schemes or (b) use large to medium-sized patches and rely on additional mechanisms within the model to capture the spatial hierarchy of details. We propose Symmetrical Hierarchical Forest (SHF), a lightweight approach that adaptively patches the input image to increase token information density and encode hierarchical spatial structures into the input embedding. We then apply a reverse depatching scheme to the output embeddings of the transformer encoder, eliminating the need for convolution-based decoders. Unlike previous methods that modify attention mechanisms \wahib{or use a complex hierarchy of interacting models}, SHF can be retrofitted to any ViT model to allow it to learn the hierarchical structure of details in high-resolution images without requiring architectural changes. Experimental results demonstrate significant gains in computational efficiency and performance: on the PAIP WSI dataset, we achieved a 3$\sim$32$\times$ speedup or a 2.95\% to 7.03\% increase in accuracy (measured by Dice score) at a $64K^2$ resolution with the same computational budget, compared to state-of-the-art production models. On the 3D medical datasets BTCV and KiTS, training was 6$\times$ faster, with accuracy gains of 6.93\% and 5.9\%, respectively, compared to models without SHF.