NeurIPS 2025 论文列表

Vision-Language Models (VLMs) have shown remarkable progress in visual understanding in recent years. Yet, they still lag behind human capabilities in specific visual tasks such as counting or relational reasoning. To understand the underlying limitations, we adopt methodologies from cognitive science, analyzing VLM performance along core cognitive axes: Perception, Attention, and Memory. Using a suite of tasks targeting these abilities, we evaluate state-of-the-art VLMs, including GPT-4o. Our analysis reveals distinct cognitive profiles: while advanced models approach ceiling performance on some tasks (e.g. category identification), a significant gap persists, particularly in tasks requiring spatial understanding or selective attention. Investigating the source of these failures and potential methods for improvement, we employ a vision-text decoupling analysis, finding that models struggling with direct visual reasoning show marked improvement when reasoning over their own generated text captions. These experiments reveal a strong need for improved VLM Chain-of-Thought (CoT) abilities, even in models that consistently exceed human performance. Furthermore, we demonstrate the potential of targeted fine-tuning on composite visual reasoning tasks and show that fine-tuning smaller VLMs moderately improves core cognitive abilities. While this improvement does not translate to large enhancements on challenging, out-of-distribution benchmarks, we show broadly that VLM performance on our datasets strongly correlates with performance on established benchmarks like MMMU-Pro and VQAv2. Our work provides a detailed analysis of VLM cognitive strengths and weaknesses and identifies key bottlenecks in simultaneous perception and reasoning while also providing an effective and simple solution.

Large vision-language models (LVLMs) are increasingly deployed in interactive applications such as virtual and augmented reality, where a first-person (egocentric) view captured by head-mounted cameras serves as key input. While this view offers fine-grained cues about user attention and hand-object interactions, its narrow field of view and lack of global context often lead to failures on spatially or contextually demanding queries. To address this, we introduce a framework that augments egocentric inputs with third-person (exocentric) views, providing complementary information such as global scene layout and object visibility to LVLMs. We present E3VQA, the first benchmark for multi-view question answering with 4K high-quality question-answer pairs grounded in synchronized ego-exo image pairs. Additionally, we propose M3CoT, a training-free prompting technique that constructs a unified scene representation by integrating scene graphs from three complementary perspectives. M3CoT enables LVLMs to reason more effectively across views, yielding consistent performance gains (4.84\% for GPT-4o and 5.94\% for Gemini 2.0 Flash) over a recent CoT baseline. Our extensive evaluation reveals key strengths and limitations of LVLMs in multi-view reasoning and highlights the value of leveraging both egocentric and exocentric inputs. The dataset and source code are available at [https://github.com/Leeinsu1/Towards-Comprehensive-Scene-Understanding](https://github.com/Leeinsu1/Towards-Comprehensive-Scene-Understanding).

Large-scale machine learning models deliver strong performance across a wide range of tasks but come with significant computational and resource constraints. To mitigate these challenges, local smaller models are often deployed alongside larger models, relying on routing and deferral mechanisms to offload complex tasks. However, existing approaches inadequately balance the capabilities of these models, often resulting in unnecessary deferrals or sub-optimal resource usage. In this work, we introduce a novel loss function called Gatekeeper for calibrating smaller models in cascade setups. Our approach fine-tunes the smaller model to confidently handle tasks it can perform correctly while deferring complex tasks to the larger model. Moreover, it incorporates a mechanism for managing the trade-off between model performance and deferral accuracy and is broadly applicable across various tasks and domains without any architectural changes. We evaluated our method on encoder-only, decoder-only, and encoder-decoder architectures. Experiments across image classification, language modeling, and vision-language tasks show that our approach substantially improves deferral performance.

Speculative decoding has become a widely adopted as an effective technique for lossless inference acceleration when deploying large language models (LLMs). While on-the-fly self-speculative methods offer seamless integration and broad utility, they often fall short of the speed gains achieved by methods relying on specialized training. Cascading a hierarchy of draft models promises further acceleration and flexibility, but the high cost of training multiple models has limited its practical application. In this paper, we propose a novel Cascade Adaptive Self-Speculative Decoding (CAS-Spec) method which constructs speculative draft models by leveraging dynamically switchable inference acceleration (DSIA) strategies, including layer sparsity and activation quantization. Furthermore, traditional vertical and horizontal cascade algorithms are inefficient when applied to self-speculative decoding methods. We introduce a Dynamic Tree Cascade (DyTC) algorithm that adaptively routes the multi-level draft models and assigns the draft lengths, based on the heuristics of acceptance rates and latency prediction. Our CAS-Spec method achieves state-of-the-art acceleration compared to existing on-the-fly speculative decoding methods, with an average speedup from $1.1\times$ to $2.3\times$ over autoregressive decoding across various LLMs and datasets. DyTC improves the average speedup by $47$\% and $48$\% over cascade-based baseline and tree-based baseline algorithms, respectively. CAS-Spec can be easily integrated into most existing LLMs and holds promising potential for further acceleration as self-speculative decoding techniques continue to evolve.

Pre-trained Vision-Language Models (VLMs) exhibit strong generalization capabilities, enabling them to recognize a wide range of objects across diverse domains without additional training. However, they often retain irrelevant information beyond the requirements of specific target downstream tasks, raising concerns about computational efficiency and potential information leakage. This has motivated growing interest in approximate unlearning, which aims to selectively remove unnecessary knowledge while preserving overall model performance. Existing approaches to approximate unlearning have primarily focused on {\em class unlearning}, where a VLM is retrained to fail to recognize specified object classes while maintaining accuracy for others. However, merely forgetting object classes is often insufficient in practical applications. For instance, an autonomous driving system should accurately recognize {\em real} cars, while avoiding misrecognition of {\em illustrated} cars depicted in roadside advertisements as {\em real} cars, which could be hazardous. In this paper, we introduce {\em Approximate Domain Unlearning (ADU)}, a novel problem setting that requires reducing recognition accuracy for images from specified domains (e.g., {\em illustration}) while preserving accuracy for other domains (e.g., {\em real}). ADU presents new technical challenges: due to the strong domain generalization capability of pre-trained VLMs, domain distributions are highly entangled in the feature space, making naive approaches based on penalizing target domains ineffective. To tackle this limitation, we propose a novel approach that explicitly disentangles domain distributions and adaptively captures instance-specific domain information. Extensive experiments on four multi-domain benchmark datasets demonstrate that our approach significantly outperforms strong baselines built upon state-of-the-art VLM tuning techniques, paving the way for practical and fine-grained unlearning in VLMs. Code : https://kodaikawamura.github.io/Domain_Unlearning/.

Action recognition is a fundamental task in video understanding. Existing methods typically extract unified features to process all actions in one video, which makes it challenging to model the interactions between different objects in multi-action scenarios. To alleviate this issue, we explore disentangling any specified actions from complex scenes as an effective solution. In this paper, we propose Prompt-guided Disentangled Representation for Action Recognition (ProDA), a novel framework that disentangles any specified actions from a multi-action scene. ProDA leverages Spatio-temporal Scene Graphs (SSGs) and introduces Dynamic Prompt Module (DPM) to guide a Graph Parsing Neural Network (GPNN) in generating action-specific representations. Furthermore, we design a video-adapted GPNN that aggregates information using dynamic weights. Extensive experiments on two complex video action datasets, Charades and SportsHHI, demonstrate the effectiveness of our approach against state-of-the-art methods. Our code can be found in https://github.com/iamsnaping/ProDA.git.

Dataset bias, where data points are skewed to certain concepts, is ubiquitous in machine learning datasets. Yet, systematically identifying these biases is challenging without costly, fine-grained attribute annotations. We present ConceptScope, a scalable and automated framework for analyzing visual datasets by discovering and quantifying human-interpretable concepts using Sparse Autoencoders trained on representations from vision foundation models. ConceptScope categorizes concepts into target, context, and bias types based on their semantic relevance and statistical correlation to class labels, enabling class-level dataset characterization, bias identification, and robustness evaluation through concept-based subgrouping. We validate that ConceptScope captures a wide range of visual concepts, including objects, textures, backgrounds, facial attributes, emotions, and actions, through comparisons with annotated datasets. Furthermore, we show that concept activations produce spatial attributions that align with semantically meaningful image regions. ConceptScope reliably detects known biases (e.g., background bias in Waterbirds) and uncovers previously unannotated ones (e.g, co-occurring objects in ImageNet), offering a practical tool for dataset auditing and model diagnostics.

DNA language models have revolutionized our ability to design and manipulate DNA sequences—the fundamental language of life—with unprecedented precision, enabling transformative applications in therapeutics, synthetic biology, and gene editing. However, this capability also poses significant dual-use risks, including the potential creation of harmful biological agents. To address these biosecurity challenges, we introduce two innovative watermarking techniques: DNAMark and CentralMark. DNAMark employs synonymous codon substitutions to embed robust watermarks in DNA sequences while preserving the function of encoded proteins. CentralMark advances this by creating inheritable watermarks that transfer from DNA to translated proteins, leveraging protein embeddings to ensure detection across the central dogma. Both methods utilize state-of-the-art embeddings to generate watermark logits, enhancing resilience against natural mutations, synthesis errors, and adversarial attacks. Evaluated on a therapeutic DNA benchmark, DNAMark and CentralMark achieve F1 detection scores above 0.85 under diverse conditions, while maintaining over 60\% sequence similarity to ground truth and degeneracy scores below 15\%. A case study on a CRISPR-Cas9 system underscores CentralMark’s utility in real-world synthetic biology. This work establishes a vital framework for securing DNA language models, balancing innovation with accountability to mitigate biosecurity risks.

A recent study has proposed the ``Master Key Filters Hypothesis" for convolutional neural network filters. This paper extends this hypothesis by radically constraining its scope to a single set of just 8 universal filters that depthwise separable convolutional networks inherently converge to. While conventional DS-CNNs employ thousands of distinct trained filters, our analysis reveals these filters are predominantly linear shifts (ax+b) of our discovered universal set. Through systematic unsupervised search, we extracted these fundamental patterns across different architectures and datasets. Remarkably, networks initialized with these 8 unique frozen filters achieve over 80\% ImageNet accuracy, and even outperform models with thousands of trainable parameters when applied to smaller datasets. The identified master key filters closely match Difference of Gaussians (DoGs), Gaussians, and their derivatives, structures that are not only fundamental to classical image processing but also strikingly similar to receptive fields in mammalian visual systems. Our findings provide compelling evidence that depthwise convolutional layers naturally gravitate toward this fundamental set of spatial operators regardless of task or architecture. This work offers new insights for understanding generalization and transfer learning through the universal language of these master key filters.

We introduce Filter Like You Test (FLYT), an algorithm for curating large-scale vision-language datasets that *learns* the usefulness of each data point as a pretraining example. FLYT trains a scoring model that learns to weigh each example's features using gradient signals from downstream tasks training sets. Based on FLYT, we implement Mixing-FLYT (M-FLYT), which takes the per-example scores generated by different scoring methods as features, and learns to unify them into a single score. FLYT naturally produces a distribution over the training examples, which we leverage through Soft Cap Sampling (SCS), a strategy for obtaining a filtered pretraining dataset from per-example probabilities that samples examples while preventing over-representation through a repetition penalty. Using these methods, we achieve 40.1\% ImageNet zero-shot accuracy on the DataComp medium scale filtering benchmark, a 2\% absolute accuracy increase over all previous results and a 5.5\% increase over results that---like us---use only public resources. Our approach also yields 37.7\% on the average of 38 DataComp evaluation tasks, outperforming previous public-resource approaches by 0.4\%.

Unified vision-language models have made significant progress in multimodal understanding and generation, yet they largely fall short in producing multimodal interleaved outputs, which is a crucial capability for tasks like visual storytelling and step-by-step visual reasoning. In this work, we propose a reinforcement learning-based post-training strategy to unlock this capability in existing unified models, without relying on large-scale multimodal interleaved datasets. We begin with a warm-up stage using a hybrid dataset comprising curated interleaved sequences and limited data for multimodal understanding and text-to-image generation, which exposes the model to interleaved generation patterns while preserving its pretrained capabilities. To further refine interleaved generation, we propose a unified policy optimization framework that extends Group Relative Policy Optimization (GRPO) to the multimodal setting. Our approach jointly models text and image generation within a single decoding trajectory and optimizes it with our novel hybrid rewards covering textual relevance, visual-text alignment, and structural fidelity. Additionally, we incorporate process-level rewards to provide step-wise guidance, enhancing training efficiency in complex multimodal tasks. Experiments on MMIE and InterleavedBench demonstrate that our approach significantly enhances the quality and coherence of multimodal interleaved generation.

Predicting human eye movements during goal-directed visual search is critical for enhancing interactive AI systems. In medical imaging, such prediction can support radiologists in interpreting complex data, such as chest X-rays. Many existing methods rely on generic vision--language models and saliency-based features, which can limit their ability to capture fine-grained clinical semantics and integrate domain knowledge effectively. We present \textbf{LogitGaze-Med}, a state-of-the-art multimodal transformer framework that unifies (1) domain-specific visual encoders (e.g., CheXNet), (2) textual embeddings of diagnostic labels, and (3) semantic priors extracted via the logit-lens from an instruction-tuned medical vision--language model (LLaVA-Med). By directly predicting continuous fixation coordinates and dwell durations, our model generates clinically meaningful scanpaths. Experiments on the GazeSearch dataset and synthetic scanpaths generated from MIMIC-CXR and validated by experts demonstrate that LogitGaze-Med improves scanpath similarity metrics by 20--30\% over competitive baselines and yields over 5\% gains in downstream pathology classification when incorporating predicted fixations as additional training data.

There is growing interest in using machine learning (ML) to support clinical diagnosis, but most approaches rely on static, fully observed datasets and fail to reflect the sequential, resource-aware reasoning clinicians use in practice. Diagnosis remains complex and error prone, especially in high-pressure or resource-limited settings, underscoring the need for frameworks that help clinicians make timely and cost-effective decisions. We propose ACTMED (Adaptive Clinical Test selection via Model-based Experimental Design), a diagnostic framework that integrates Bayesian Experimental Design (BED) with large language models (LLMs) to better emulate real-world diagnostic reasoning. At each step, ACTMED selects the test expected to yield the greatest reduction in diagnostic uncertainty for a given patient. LLMs act as flexible simulators, generating plausible patient state distributions and supporting belief updates without requiring structured, task-specific training data. Clinicians can remain in the loop; reviewing test suggestions, interpreting intermediate outputs, and applying clinical judgment throughout. We evaluate ACTMED on real-world datasets and show it can optimize test selection to improve diagnostic accuracy, interpretability, and resource use. This represents a step toward transparent, adaptive, and clinician-aligned diagnostic systems that generalize across settings with reduced reliance on domain-specific data.

Modifying well-trained models for purposes such as pruning or unlearning, without access to training data or the original loss function, is a challenging problem. While techniques exist for such modification, they often require training data, are computationally expensive, or are architecture-specific. To address this, we investigate the fundamental question of identifying components that are critical to the model’s predictive performance, without access to either gradients or the loss function, and with only distributional access such as synthetic data. We theoretically demonstrate that the global reconstruction error is linearly bounded by local reconstruction errors for Lipschitz-continuous networks such as CNNs and well-trained Transformers (which, contrary to existing literature, we find exhibit Lipschitz continuity). This motivates using the locally reconstructive behavior of component subsets to quantify their global importance, via a metric that we term *Subset Fidelity*. In the uncorrelated features setting, selecting individual components via their Subset Fidelity scores is optimal, which we use to propose **ModHiFi**, an algorithm for model modification that requires no training data or loss function access. **ModHiFi-P**, for structured pruning, achieves an 11% speedup over the current state of the art on ImageNet models and competitive performance on language models. **ModHiFi-U**, for classwise unlearning, achieves complete unlearning on CIFAR-10 without fine-tuning and demonstrates competitive performance on Swin Transformers.

Grounding large language models (LLMs) in domain-specific tasks like post-hoc dash-cam driving video analysis is challenging due to their general-purpose training and lack of structured inductive biases. As vision is often the sole modality available for such analysis (i.e., no LiDAR, GPS, etc.), existing video-based vision-language models (V-VLMs) struggle with spatial reasoning, causal inference, and explainability of events in the input video. To this end, we introduce iFinder, a structured semantic grounding framework that decouples perception from reasoning by translating dash-cam videos into a hierarchical, interpretable data structure for LLMs. iFinder operates as a modular, training-free pipeline that employs pretrained vision models to extract critical cues—object pose, lane positions, and object trajectories—which are hierarchically organized into frame- and video-level structures. Combined with a three-block prompting strategy, it enables step-wise, grounded reasoning for the LLM to refine a peer V-VLM's outputs and provide accurate reasoning. Evaluations on four public dash-cam video benchmarks show that iFinder's proposed grounding with domain-specific cues—especially object orientation and global context—significantly outperforms end-to-end V-VLMs on four zero-shot driving benchmarks, with up to 39% gains in accident reasoning accuracy. By grounding LLMs with driving domain-specific representations, iFinder offers a zero-shot, interpretable, and reliable alternative to end-to-end V-VLMs for post-hoc driving video understanding.

In this paper, we introduce knowledge image generation as a new task, alongside the Massive Multi-Discipline Multi-Tier Knowledge-Image Generation Benchmark (MMMG) to probe the reasoning capability of image generation models. Knowledge images have been central to human civilization and to the mechanisms of human learning—a fact underscored by dual-coding theory and the picture-superiority effect. Generating such images is challenging, demanding multimodal reasoning that fuses world knowledge with pixel-level grounding into clear explanatory visuals. To enable comprehensive evaluation, MMMG offers $4,456$ expert-validated (knowledge) image-prompt pairs spanning $10$ disciplines, $6$ educational levels, and diverse knowledge formats such as charts, diagrams, and mind maps. To eliminate confounding complexity during evaluation, we adopt a unified Knowledge Graph (KG) representation. Each KG explicitly delineates a target image’s core entities and their dependencies. We further introduce MMMG-Score‌ to evaluate generated knowledge images. This metric combines factual fidelity, measured by graph-edit distance between KGs, with visual clarity assessment. Comprehensive evaluations of $21$ state-of-the-art text-to-image generation models expose serious reasoning deficits—low entity fidelity, weak relations, and clutter—with GPT-4o achieving an MMMG-Score‌ of only $50.20$, underscoring the benchmark’s difficulty. To spur further progress, we release FLUX-Reason (MMMG-Score‌ of $34.45$), an effective and open baseline that combines a reasoning LLM with diffusion models and is trained on $16,000$ curated knowledge image–prompt pairs.

To migrate the remarkable successes of Large Language Models (LLMs), the community has made numerous efforts to generalize them to the table reasoning tasks for the widely deployed tabular data. Despite that, in this work, by showing a probing experiment on our proposed StructQA benchmark, we postulate that even the most advanced LLMs (such as GPTs) may still fall short of coping with tabular data. More specifically, the current scheme often simply relies on serializing the tabular data, together with the meta information, then inputting them through the LLMs. We argue that the loss of structural information is the root of this shortcoming. In this work, we further propose TAMO, which bears an ideology to treat the tables as an independent modality integrated with the text tokens. The resulting model in TAMO is a multimodal framework consisting of a hypergraph neural network as the global table encoder seamlessly integrated with the mainstream LLM. Empirical results on various benchmarking datasets, including HiTab, WikiTQ, WikiSQL, FeTaQA, and StructQA, have demonstrated significant improvements on generalization with an average relative gain of **42.65%**.

The task of localizing an object's spatial tube based on language instructions and video, known as spatial video grounding (SVG), has attracted widespread interest. Existing SVG tasks have focused on ego-centric fixed front perspective and simple scenes, which only involved a very limited view and environment. However, UAV-based SVG remains underexplored, which neglects the inherent disparities in drone movement and the complexity of aerial object localization. To facilitate research in this field, we introduce the novel spatial aerial video grounding (SAVG) task. Specifically, we meticulously construct a large-scale benchmark, UAV-SVG, which contains over 2 million frames and offers 216 highly diverse target categories. To address the disparities and challenges posed by complex aerial environments, we propose a new end-to-end transformer architecture, coined SAVG-DETR. The innovations are three-fold. 1) To overcome the computational explosion of self-attention when introducing multi-scale features, our encoder efficiently decouples the multi-modality and multi-scale spatio-temporal modeling into intra-scale multi-modality interaction and cross-scale visual-only fusion. 2) To enhance small object grounding ability, we propose the language modulation module to integrate multi-scale information into language features and the multi-level progressive spatial decoder to decode from high to low level. The decoding stage for the lower-level vision-language features is gradually increased. 3) To improve the prediction consistency across frames, we design the decoding paradigm based on offset generation. At each decoding stage, we utilize reference anchors to constrict the grounding region, use context-rich object queries to predict offsets, and update reference anchors for the next stage. From coarse to fine, our SAVG-DETR gradually bridges the modality gap and iteratively refines reference anchors of the referred object, eventually grounding the spatial tube. Extensive experiments demonstrate that our SAVG-DETR significantly outperforms existing state-of-the-art methods. The dataset and code will be available at here.

Multimodal large language models (MLLMs) remain vulnerable to transferable adversarial examples. While existing methods typically achieve targeted attacks by aligning global features—such as CLIP’s [CLS] token—between adversarial and target samples, they often overlook the rich local information encoded in patch tokens. This leads to suboptimal alignment and limited transferability, particularly for closed-source models. To address this limitation, we propose a targeted transferable adversarial attack method based on feature optimal alignment, called FOA-Attack, to improve adversarial transfer capability. Specifically, at the global level, we introduce a global feature loss based on cosine similarity to align the coarse-grained features of adversarial samples with those of target samples. At the local level, given the rich local representations within Transformers, we leverage clustering techniques to extract compact local patterns to alleviate redundant local features. We then formulate local feature alignment between adversarial and target samples as an optimal transport (OT) problem and propose a local clustering optimal transport loss to refine fine-grained feature alignment. Additionally, we propose a dynamic ensemble model weighting strategy to adaptively balance the influence of multiple models during adversarial example generation, thereby further improving transferability. Extensive experiments across various models demonstrate the superiority of the proposed method, outperforming state-of-the-art methods, especially in transferring to closed-source MLLMs.