NeurIPS 2025 论文列表

We introduce WearVQA, the first benchmark specifically designed to evaluate the visual question answering (VQA) capabilities of multi-modal AI assistant on wearable devices like smart glasses. Unlike prior benchmarks that focus on high-quality, third-person imagery, WearVQA reflects the unique chal- lenges of ego-centric interaction—where visual inputs may be occluded, poorly lit, unzoomed, or blurry, and questions are grounded in realistic wearable use cases. The benchmark comprises 2,500 carefully curated image-question-answer triplets, spanning 7 diverse image domains including both text-centric and general scenes, 10 cognitive task types ranging from basic recognition to various forms of reasoning, and 6 common wearables-specific image quality issues. All questions are designed to be answerable using only the visual input and common senses. WearVQA is paired with a rigorous LLM-as-a-judge evaluation framework with 96% labeling accuracy. Open-source and proprietary multi-modal LLMs achieved a QA accuracy as low as 24–52% on WearVQA, with substantial drops on lower-quality images and reasoning- heavy tasks. These observations position WearVQA as a comprehensive and challenging benchmark for guiding technicial advancement towards robust, real-world multi-modal wearables AI systems.

Humans excel at constructing panoramic mental models of their surroundings, maintaining object permanence and inferring scene structure beyond visible regions. In contrast, current artificial vision systems struggle with persistent, panoramic understanding, often processing scenes egocentrically on a frame-by-frame basis. This limitation is pronounced in the Track Any Point (TAP) task, where existing methods fail to track 2D points outside the field of view. To address this, we introduce TAP-Vid 360, a novel task that requires predicting the 3D direction to queried scene points across a video sequence, even when far outside the narrow field of view of the observed video. This task fosters learning allocentric scene representations without needing dynamic 4D ground truth scene models for training. Instead, we exploit 360 videos as a source of supervision, resampling them into narrow field-of-view perspectives while computing ground truth directions by tracking points across the full panorama using a 2D pipeline. We introduce a new dataset and benchmark, TAP360-10k comprising 10k perspective videos with ground truth directional point tracking. Our baseline adapts CoTracker v3 to predict per-point rotations for direction updates, outperforming existing TAP and TAP-Vid 3D methods.

Person retrieval has attracted rising attention. Existing methods are mainly divided into two retrieval modes, namely image-only and text-only. However, they are unable to make full use of the available information and are difficult to meet diverse application requirements. To address the above limitations, we propose a new Composed Person Retrieval (CPR) task, which combines visual and textual queries to identify individuals of interest from large-scale person image databases. Nevertheless, the foremost difficulty of the CPR task is the lack of available annotated datasets. Therefore, we first introduce a scalable automatic data synthesis pipeline, which decomposes complex multimodal data generation into the creation of textual quadruples followed by identity-consistent image synthesis using fine-tuned generative models. Meanwhile, a multimodal filtering method is designed to ensure the resulting SynCPR dataset retains 1.15 million high-quality and fully synthetic triplets. Additionally, to improve the representation of composed person queries, we propose a novel Fine-grained Adaptive Feature Alignment (FAFA) framework through fine-grained dynamic alignment and masked feature reasoning. Moreover, for objective evaluation, we manually annotate the Image-Text Composed Person Retrieval (ITCPR) test set. The extensive experiments demonstrate the effectiveness of the SynCPR dataset and the superiority of the proposed FAFA framework when compared with the state-of-the-art methods. All code and data will be provided at https://github.com/Delong-liu-bupt/Composed_Person_Retrieval.

Model merging provides a scalable alternative to multi-task training by combining specialized finetuned models through parameter arithmetic, enabling efficient deployment without the need for joint training or access to all task data. While recent methods have shown promise, existing evaluations are limited in both model scale and task diversity, leaving open questions about their applicability to large, domain-specialized LLMs. To tackle the challenges, we introduce MergeBench, a comprehensive evaluation suite designed to assess model merging at scale. MergeBench builds on state-of-the-art open-source language models, including Llama and Gemma families at 2B to 9B scales, and covers five key domains: instruction following, mathematics, multilingual understanding, coding and safety. We standardize finetuning and evaluation protocols, and assess eight representative merging methods across multi-task performance, forgetting and runtime efficiency. Based on extensive experiments, we provide practical guidelines for algorithm selection and share insights showing that model merging tends to perform better on stronger base models, with techniques such as merging coefficient tuning and sparsification improving knowledge retention. However, several challenges remain, including the computational cost on large models, the gap for in-domain performance compared to multi-task models, and the underexplored role of model merging in standard LLM training pipelines. We hope MergeBench provides a foundation for future research to advance the understanding and practical application of model merging.

We propose a novel AutoRegressive Generation-based paradigm for image Segmentation (ARGenSeg), achieving multimodal understanding and pixel-level perception within a unified framework. Prior works integrating image segmentation into multimodal large language models (MLLMs) typically employ either boundary points representation or dedicated segmentation heads. These methods rely on discrete representations or semantic prompts fed into task-specific decoders, which limits the ability of the MLLM to capture fine-grained visual details. To address these challenges, we introduce a segmentation framework for MLLM based on image generation, which naturally produces dense masks for target objects. We leverage MLLM to output visual tokens and detokenize them into images using an universal VQ-VAE, making the segmentation fully dependent on the pixel-level understanding of the MLLM. To reduce inference latency, we employ a next-scale-prediction strategy to generate required visual tokens in parallel. Extensive experiments demonstrate that our method surpasses prior state-of-the-art approaches on multiple segmentation datasets with a remarkable boost in inference speed, while maintaining strong understanding capabilities.

Unlocking spatial reasoning in Multimodal Large Language Models (MLLMs) is crucial for enabling intelligent interaction with 3D environments. While prior efforts often rely on explicit 3D inputs or specialized model architectures, we ask: can MLLMs reason about 3D space using only structured 2D representations derived from perception? In this work, we introduce Struct2D, a perception-guided prompting framework that combines bird’s-eye-view (BEV) images with object marks and object-centric metadata, optionally incorporating egocentric keyframes when needed. Using Struct2D, we conduct an in-depth zero-shot analysis of closed-source MLLMs (e.g., GPT-o3) and find that they exhibit surprisingly strong spatial reasoning abilities when provided with projected 2D inputs, effectively handling tasks such as relative direction estimation and route planning. Motivated by these findings, we construct a large-scale instructional tuning dataset, \textbf{Struct2D-Set}, using an automated pipeline that generates fine-grained QA pairs grounded in 3D indoor scenes. We then fine-tune an open-source MLLM (Qwen2.5VL) using Struct2D-Set, relying on noisy 3D perception rather than ground-truth annotations. Despite this, the tuned model achieves strong performance across multiple spatial reasoning benchmarks, including 3D question answering, captioning, and object grounding, spanning eight diverse reasoning categories. Our approach demonstrates that structured 2D inputs can effectively bridge perception and language reasoning in MLLMs—without requiring explicit 3D representations as input. We will release both our code and dataset to support future research.

Color plays an important role in human perception and usually provides critical clues in visual reasoning. However, it is unclear whether and how vision-language models (VLMs) can perceive, understand, and leverage color as humans. This paper introduces ColorBench, an innovative benchmark meticulously crafted to assess the capabilities of VLMs in color understanding, including color perception, reasoning, and robustness. By curating a suite of diverse test scenarios, with grounding in real applications, ColorBench evaluates how these models perceive colors, infer meanings from color-based cues, and maintain consistent performance under varying color transformations. Through an extensive evaluation of 32 VLMs with varying language models and vision encoders, our paper reveals some undiscovered findings: (i) The scaling law (larger models are better) still holds on ColorBench, while the language model plays a more important role than the vision encoder. (ii) However, the performance gaps across models are relatively small, indicating that color understanding has been largely neglected by existing VLMs. (iii) CoT reasoning improves color understanding accuracies and robustness, though they are vision-centric tasks. (iv) Color clues are indeed leveraged by VLMs on ColorBench but they can also mislead models in some tasks. These findings highlight the critical limitations of current VLMs and underscore the need to enhance color comprehension. Our ColorBench can serve as a foundational tool for advancing the study of human-level color understanding of multimodal AI.

Test-time adaptation (TTA) enhances the zero-shot robustness under distribution shifts by leveraging unlabeled test data during inference. Despite notable advances, several challenges still limit its broader applicability. First, most methods rely on backpropagation or iterative optimization, which limits scalability and hinders real-time deployment. Second, they lack explicit modeling of class-conditional feature distributions. This modeling is crucial for producing reliable decision boundaries and calibrated predictions, but it remains underexplored due to the lack of both source data and supervision at test time. In this paper, we propose ADAPT, an Advanced Distribution-Aware and backPropagation-free Test-time adaptation method. We reframe TTA as a probabilistic inference task by modeling class-conditional likelihoods using gradually updated class means and a shared covariance matrix. This enables closed-form, training-free inference. To correct potential likelihood bias, we introduce lightweight regularization guided by CLIP priors and a historical knowledge bank. ADAPT requires no source data, no gradient updates, and no full access to target data, supporting both online and transductive settings. Extensive experiments across diverse benchmarks demonstrate that our method achieves state-of-the-art performance under a wide range of distribution shifts with superior scalability and robustness.

We present the B-spline Encoded Action Sequence Tokenizer (BEAST), a novel action tokenizer that encodes action sequences into compact discrete or continuous tokens using B-splines. In contrast to existing action tokenizers based on vector quantization or byte pair encoding, BEAST requires no separate tokenizer training and consistently produces tokens of uniform length, enabling fast action sequence generation via parallel decoding. Leveraging our B-spline formulation, BEAST inherently ensures generating smooth trajectories without discontinuities between adjacent segments. We extensively evaluate BEAST by integrating it with three distinct model architectures: a Variational Autoencoder (VAE) with continuous tokens, a decoder-only Transformer with discrete tokens, and Florence-2, a pretrained Vision-Language Model with an encoder-decoder architecture, demonstrating BEAST's compatibility and scalability with large pretrained models. We evaluate BEAST across three established benchmarks consisting of 166 simulated tasks and on three distinct robot settings with a total of 8 real-world tasks. Experimental results demonstrate that BEAST (i) significantly reduces both training and inference computational costs, and (ii) consistently generates smooth, high-frequency control signals suitable for continuous control tasks while (iii) reliably achieves competitive task success rates compared to state-of-the-art methods.

Numerous well-established studies have demonstrated the superhuman capabilities of modern Vision-Language Models (VLMs) across a wide range of tasks. However, growing is the doubt about the continuing availability of reliable high-quality labeling (supervision) from human annotators, leading to stagnation of the model's performance. To address this challenge, ``superalignment'' employs the so-called weak-to-strong generalization paradigm, where the supervision from a weak model can provide generalizable knowledge for a strong model. While effective in aligning knowledge for clean samples between the strong and weak models, the standard weak-to-strong approach typically fails to capture adversarial robustness, exposing strong VLMs to adversarial attacks. This inability to transfer adversarial robustness is because adversarial samples are normally missing in the superalignment stage. To this end, we are the first to propose the weak-to-strong (adversarial) robustness generalization method to elicit zero-shot robustness in large-scale models by an unsupervised scheme, mitigating the unreliable information source for alignment from two perspectives: alignment re-weighting and source guidance refinement. We analyze settings under which robustness generalization is possible. Extensive experiments across various vision-language benchmarks validate the effectiveness of our method in numerous scenarios, demonstrating its plug-and-play applicability to large-scale VLMs.

The expanding computational costs and limited resources underscore the critical need for budgeted-iteration training, which aims to achieve optimal learning within predetermined iteration budgets. While learning rate schedules fundamentally govern the performance of different networks and tasks, particularly in budgeted-iteration scenarios, their design remains largely heuristic, lacking theoretical foundations. In addition, the optimal learning rate schedule requires extensive trial-and-error selection, making the training process inefficient. In this work, we propose the Unified Budget-Aware (UBA) schedule, a theoretically grounded learning rate schedule that consistently outperforms commonly-used schedules among diverse architectures and tasks under different constrained training budgets. First, we bridge the gap by constructing a novel training budget-aware optimization framework, which explicitly accounts for the robustness to landscape curvature variations. From this framework, we derive the UBA schedule, controlled by a single hyper-parameter $\varphi$ that provides a trade-off between flexibility and simplicity, eliminating the need for per-network numerical optimization. Moreover, we establish a theoretical connection between $\varphi$ and the condition number, adding interpretation and justification to our approach. Besides, we prove the convergence for different values of $\varphi$. We offer practical guidelines for $\varphi$ selection via theoretical analysis and empirical results. Extensive experimental results show that UBA $\textit{consistently surpasses}$ the commonly-used schedules across diverse vision and language tasks, spanning network architectures (e.g., ResNet, OLMo) and scales, under different training-iteration budgets.

Multimodal federated learning in real-world settings often encounters incomplete and heterogeneous data across clients. This results in misaligned local feature representations that limit the effectiveness of model aggregation. Unlike prior work that assumes either differing modality sets without missing input features or a shared modality set with missing features across clients, we consider a more general and realistic setting where each client observes a different subset of modalities and might also have missing input features within each modality. To address the resulting misalignment in learned representations, we propose a new federated learning framework featuring locally adaptive representations based on learnable client-side embedding controls that encode each client’s data-missing patterns. These embeddings serve as reconfiguration signals that align the globally aggregated representation with each client's local context, enabling more effective use of shared information. Furthermore, the embedding controls can be algorithmically aggregated across clients with similar data-missing patterns to enhance the robustness of reconfiguration signals in adapting the global representation. Empirical results on multiple federated multimodal benchmarks with diverse data-missing patterns across clients demonstrate the efficacy of the proposed method, achieving up to 36.45\% performance improvement under severe data incompleteness. The method is also supported by a theoretical analysis with an explicit performance bound that matches our empirical observations. Our source codes are provided at [https://github.com/nmduonggg/PEPSY](https://github.com/nmduonggg/PEPSY)

This paper firstly addresses the challenge of few-shot universal cross-domain retrieval (FS-UCDR), enabling machines trained with limited data to generalize to novel retrieval scenarios, with queries from entirely unknown domains and categories. To achieve this, we first formally define the FS-UCDR task and propose the Multi-Modal Interactive Agent Layer (MAIL), which enhances the cross-modal interaction in vision-language models (VLMs) by aligning the parameter updates of target layer pairs across modalities. Specifically, MAIL freezes the selected target layer pair and introduces a trainable agent layer pair to approximate localized parameter updates. A bridge function is then introduced to couple the agent layer pair, enabling gradient communication across modalities to facilitate update alignment. The proposed MAIL offers four key advantages: 1) its cross-modal interaction mechanism improves knowledge acquisition from limited data, making it highly effective in low-data scenarios; 2) during inference, MAIL integrates seamlessly into the VLM via reparameterization, preserving inference complexity; 3) extensive experiments validate the superiority of MAIL, which achieves substantial performance gains over data-efficient UCDR methods while requiring significantly fewer training samples; 4) beyond UCDR, MAIL also performs competitively on few-shot classification tasks, underscoring its strong generalization ability. Code.

Reinforcement learning (RL) has shown promise in enhancing the general Chain-of-Thought (CoT) reasoning capabilities of multimodal large language models (MLLMs). However, when applied to improve general CoT reasoning, existing RL frameworks often struggle to generalize beyond the training distribution. To address this, we propose NoisyGRPO, a systematic multimodal RL framework that introduces controllable noise into visual inputs for enhanced exploration and explicitly models the advantage estimation process via a Bayesian framework. Specifically, NoisyGRPO improves RL training by: (1) \textbf{Noise-Injected Exploration Policy}: Perturbing visual inputs with Gaussian noise to encourage exploration across a wider range of visual scenarios; and (2) \textbf{Bayesian Advantage Estimation}: Formulating advantage estimation as a principled Bayesian inference problem, where the injected noise level serves as a prior and the observed trajectory reward as the likelihood. This Bayesian modeling fuses both sources of information to compute a robust posterior estimate of trajectory advantage, effectively guiding MLLMs to prefer visually grounded trajectories over noisy ones. Experiments on standard CoT quality, general capability, and hallucination benchmarks demonstrate that NoisyGRPO substantially improves generalization and robustness, especially in RL settings with small-scale MLLMs such as Qwen2.5-VL 3B.

Vision-Language-Action (VLA) models have advanced robotic control by enabling end-to-end decision-making directly from multimodal inputs. However, their tightly coupled architectures expose novel security vulnerabilities. Unlike traditional adversarial perturbations, backdoor attacks represent a stealthier, persistent, and practically significant threat—particularly under the emerging Training-as-a-Service paradigm—but remain largely unexplored in the context of VLA models. To address this gap, we propose **BadVLA**, a backdoor attack method based on Objective-Decoupled Optimization, which for the first time exposes the backdoor vulnerabilities of VLA models. Specifically, it consists of a two-stage process: (1) explicit feature-space separation to isolate trigger representations from benign inputs, and (2) conditional control deviations that activate only in the presence of the trigger, while preserving clean-task performance. Empirical results on multiple VLA benchmarks demonstrate that BadVLA consistently achieves near-100\% attack success rates with minimal impact on clean task accuracy. Further analyses confirm its robustness against common input perturbations, task transfers, and model fine-tuning, underscoring critical security vulnerabilities in current VLA deployments. Our work offers the first systematic investigation of backdoor vulnerabilities in VLA models, highlighting an urgent need for secure and trustworthy embodied model design practices.

Structural pruning enhances hardware-agnostic inference efficiency for large language models (LLMs) yet often fails to maintain comparable performance. Local pruning performs efficient layer-by-layer compression but ignores global topology. Although global pruning aims to identify an optimal sparse model, intuitive methods typically adopt a two-stage paradigm that first evaluates substructure saliency and then applies global pruning, which ignores inter-structure dependencies and fails to achieve end-to-end optimization. To address these limitations, we propose Týr-the-Pruner, an efficient end-to-end search-based global structural pruning framework. This framework constructs a supernet by repeatedly applying local pruning across a range of sparsity ratios to each layer in an LLM, with the core goal of determining the optimal sparsity distribution under a target overall sparsity ratio. Concretely, we introduce an effective local pruning and an expectation error accumulation approach to improve supernet construction. Furthermore, we employ an iterative prune-and-search strategy with coarse-to-fine sparsity granularity to ensure efficient search convergence. Experimental results show that Týr-the-Pruner achieves state-of-the-art structural pruning, retaining 97% of the dense model's performance while removing a challenging 50% of Llama-3.1-70B's parameters.

Large Language Models (LLMs) are increasingly being explored across a range of reasoning tasks. However, LLMs sometimes struggle with reasoning tasks under uncertainty that are relatively easy for humans, such as planning actions in stochastic environments. The adoption of LLMs for reasoning is impeded by uncertainty challenges, such as LLM uncertainty and environmental uncertainty. LLM uncertainty arises from the stochastic sampling process inherent to LLMs. Most LLM-based Decision-Making (LDM) approaches address LLM uncertainty through multiple reasoning chains or search trees. However, these approaches overlook environmental uncertainty, which leads to poor performance in environments with stochastic state transitions. Some recent LDM approaches deal with uncertainty by forecasting the probability of unknown variables. However, they are not designed for multi-step reasoning tasks that require interaction with the environment. To address uncertainty in LLM decision-making, we introduce PlanU, an LLM-based planning method that captures uncertainty within Monte Carlo Tree Search (MCTS). PlanU models the return of each node in the MCTS as a quantile distribution, which uses a set of quantiles to represent the return distribution. To balance exploration and exploitation during tree search, PlanU introduces an Upper Confidence Bounds with Curiosity (UCC) score which estimates the uncertainty of MCTS nodes. Through extensive experiments, we demonstrate the effectiveness of PlanU in LLM-based reasoning tasks under uncertainty.

Cybersecurity and software research have crossed paths with modern deep learning research for a few years. The power of large language models (LLMs) in particular has intrigued us to apply them to understanding binary code. In this paper, we investigate some of the many ways LLMs can be applied to binary code similarity detection, as it is a significantly more difficult task compared to source code similarity detection due to the sparsity of information and less meaningful syntax. It also has great practical implications, such as vulnerability and malware detection. We find that pretrained LLMs are mostly capable of detecting similar binary code, even with a zero-shot setting. Our main contributions and findings are to provide several supervised fine-tuning methods that, when combined, significantly surpass zero-shot LLMs and state-of-the-art binary code similarity detection methods. Specifically, we up-train the model through data augmentation, translation-style causal learning, LLM2Vec, and cumulative GTE loss. With a complete ablation study, we show that our training method can transform a generic language model into a powerful binary similarity expert, and is also robust and general enough for cross-optimization, cross-architecture, and cross-obfuscation detection.

Multimodal Large Language Models (MLLMs) have shown great potential in revolutionizing Graphical User Interface (GUI) automation. However, existing GUI models mostly rely on learning from nearly error-free offline trajectories, thus lacking reflection and error recovery capabilities. To bridge this gap, we propose GUI-Reflection, a novel framework that explicitly integrates self-reflection and error correction capabilities into end-to-end multimodal GUI models throughout dedicated training stages: GUI-specific pre-training, offline supervised fine-tuning (SFT), and online reflection tuning. GUI-reflection enables self-reflection behavior emergence with fully automated data generation and learning processes without requiring any human annotation. Specifically, 1) we first propose scalable data pipelines to automatically construct reflection and error correction data from existing successful trajectories. While existing GUI models mainly focus on grounding and UI understanding ability, we propose the GUI-Reflection Task Suite to learn and evaluate reflection-oriented abilities explicitly. 2) Furthermore, we built a diverse and efficient environment for online training and data collection of GUI models on mobile devices. 3) We also present an iterative online reflection tuning algorithm leveraging the proposed environment, enabling the model to continuously enhance its reflection and error correction abilities. Our framework equips GUI agents with self-reflection and correction capabilities, paving the way for more robust, adaptable, and intelligent GUI automation, with all data, models, environments, and tools to be released publicly.