NeurIPS 2025 论文列表

Endowing robots with tool design abilities is critical for enabling them to solve complex manipulation tasks that would otherwise be intractable. While recent generative frameworks can automatically synthesize task settings—such as 3D scenes and reward functions—they have not yet addressed the challenge of tool-use scenarios. Simply retrieving human-designed tools might not be ideal since many tools (e.g., a rolling pin) are difficult for robotic manipulators to handle. Furthermore, existing tool design approaches either rely on predefined templates with limited parameter tuning or apply generic 3D generation methods that are not optimized for tool creation. To address these limitations, we propose **RobotSmith**, an automated pipeline that leverages the implicit physical knowledge embedded in vision-language models (VLMs) alongside the more accurate physics provided by physics simulations to design and use tools for robotic manipulation. Our system (1) iteratively proposes tool designs using collaborative VLM agents, (2) generates low-level robot trajectories for tool use, and (3) jointly optimizes tool geometry and usage for task performance. We evaluate our approach across a wide range of manipulation tasks involving rigid, deformable, and fluid objects. Experiments show that our method consistently outperforms strong baselines in both task success rate and overall performance. Notably, our approach achieves a 50.0\% average success rate, significantly surpassing other baselines such as 3D generation (21.4\%) and tool retrieval (11.1\%). Finally, we deploy our system in real-world settings, demonstrating that the generated tools and their usage plans transfer effectively to physical execution, validating the practicality and generalization capabilities of our approach.

Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as science, law, and healthcare, where accurate expressions of uncertainty are essential for reliability and trust. However, current LLMs are often observed to generate incorrect answers with high confidence—a phenomenon known as "overconfidence". Recent efforts have focused on calibrating LLMs' verbalized confidence: i.e., their expressions of confidence in text form, such as "I am 80% confident that...". Existing approaches either rely on prompt engineering or fine-tuning with heuristically generated uncertainty estimates, both of which have limited effectiveness and generalizability. Motivated by the notion of proper scoring rules for calibration in classical machine learning models, we introduce ConfTuner, a simple and efficient fine-tuning method that introduces minimal overhead and does not require ground-truth confidence scores or proxy confidence estimates. ConfTuner relies on a new loss function, tokenized Brier score, which we theoretically prove to be a proper scoring rule, intuitively meaning that it "correctly incentivizes the model to report its true probability of being correct". ConfTuner improves calibration across diverse reasoning tasks and generalizes to black-box models such as GPT-4o. Our results further show that better-calibrated confidence enables downstream gains in self-correction and model cascade, advancing the development of trustworthy LLM systems. The code is available at https://github.com/liushiliushi/ConfTuner.

Despite Multimodal Large Language Models (MLLMs) showing promising results on general zero-shot image classification tasks, fine-grained image classification remains challenging. It demands precise attention to subtle visual details to distinguish between visually similar subcategories—details that MLLMs may easily overlook without explicit guidance. To address this, we introduce AutoSEP, an iterative self-supervised prompt learning framework designed to enhance MLLM fine-grained classification capabilities in a fully unsupervised manner. Our core idea is to leverage unlabeled data to learn a description prompt that guides MLLMs in identifying crucial discriminative features within an image, and boost classification accuracy. We developed an automatic self-enhancing prompt learning framework called AutoSEP to iteratively improve the description prompt using unlabeled data, based on instance-level classification scoring function. AutoSEP only requires black-box access to MLLMs, eliminating the need for any training or fine-tuning. We evaluate our approach on multiple fine-grained classification datasets. It consistently outperforms other unsupervised baselines, demonstrating the effectiveness of our self-supervised optimization framework. Notably, AutoSEP in average improves 13\% over standard zero-shot classification and 3\% over the best-performing baselines.

Remote physiological sensing is an evolving area of research. As systems approach clinical precision, there is increasing focus on complex applications such as cognitive state estimation. Hence, there is a need for large datasets that facilitate research into complex downstream tasks such as remote cognitive load estimation. A first-of-its-kind, our paper introduces an open-source multimodal multi-vital sign dataset consisting of concurrent recordings from RGB, NIR (near-infrared), thermal, and RF (radio-frequency) sensors alongside contact-based physiological signals, such as pulse oximeter and chest bands, providing a benchmark for cognitive state assessment. By adopting a multimodal approach to remote health sensing, our dataset and its associated hardware system excel at modeling the complexities of cognitive load. Here, cognitive load is defined as the mental effort exerted during tasks such as reading, memorizing, and solving math problems. By using the NASA-TLX survey, we set personalized thresholds for defining high/low cognitive levels, enabling a more reliable benchmark. Our benchmarking scheme bridges the gap between existing remote sensing strategies and cognitive load estimation techniques by using vital signs (such as photoplethysmography (PPG) and respiratory waveforms) and physiological signals (blink waveforms) as an intermediary. Through this paper, we focus on replacing the need for intrusive contact-based physiological measurements with more user-friendly remote sensors. Our benchmarking demonstrates that multimodal fusion significantly improves remote vital sign estimation, with our fusion model achieving $<3~BPM$ (beats per minute) error for vital sign estimation. For cognitive load classification, the combination of remote PPG, remote respiratory signals, and blink markers achieves $86.49$% accuracy, approaching the performance of contact-based sensing ($87.5$%) and validating the feasibility of non-intrusive cognitive monitoring.

Large language models (LLMs) have gained significant attention in chemistry. However, most existing datasets center on molecular-level property prediction and overlook the role of fine-grained functional group (FG) information. Incorporating FG-level data can provide valuable prior knowledge that links molecular structures with textual descriptions, which can be used to build more interpretable, structure-aware LLMs for reasoning on molecule-related tasks. Moreover, LLMs can learn from such fine-grained information to uncover hidden relationships between specific functional groups and molecular properties, thereby advancing molecular design and drug discovery. Here, we introduce FGBench, a dataset comprising 625K molecular property reasoning problems with functional group information. Functional groups are precisely annotated and localized within the molecule, which ensures the dataset's interoperability thereby facilitating further multimodal applications. FGBench includes both regression and classification tasks on 245 different functional groups across three categories for molecular property reasoning: (1) single functional group impacts, (2) multiple functional group interactions, and (3) direct molecular comparisons. In the benchmark of state-of-the-art LLMs on 7K curated data, the results indicate that current LLMs struggle with FG-level property reasoning, highlighting the need to enhance reasoning capabilities in LLMs for chemistry tasks. We anticipate that the methodology employed in FGBench to construct datasets with functional group-level information will serve as a foundational framework for generating new question–answer pairs, enabling LLMs to better understand fine-grained molecular structure–property relationships. The dataset and evaluation code are available at this \href{https://github.com/xuanliugit/FGBench}{link}.

Medical Visual Question Answering (Med-VQA) holds significant potential for clinical decision support, yet existing efforts primarily focus on 2D imaging with limited task diversity. This paper presents 3D-RAD, a large-scale dataset designed to advance 3D Med-VQA using radiology CT scans. The 3D-RAD dataset encompasses six diverse VQA tasks: anomaly detection, image observation, medical computation, existence detection, static temporal diagnosis, and longitudinal temporal diagnosis. It supports both open- and closed-ended questions while introducing complex reasoning challenges, including computational tasks and multi-stage temporal analysis, to enable comprehensive benchmarking. Extensive evaluations demonstrate that existing vision-language models (VLMs), especially medical VLMs exhibit limited generalization, particularly in multi-temporal tasks, underscoring the challenges of real-world 3D diagnostic reasoning. To drive future advancements, we release a high-quality training set 3D-RAD-T of 136,195 expert-aligned samples, showing that fine-tuning on this dataset could significantly enhance model performance. Our dataset and code, aiming to catalyze multimodal medical AI research and establish a robust foundation for 3D medical visual understanding, are publicly available.

Few-shot learning (FSL) aims to recognize novel concepts from only a few labeled support samples. Recent studies enhance support features by incorporating additional semantic information (e.g., class descriptions) or designing complex semantic fusion modules. However, these methods still suffer from hallucinating semantics that contradict the visual evidence due to the lack of grounding in actual instances, resulting in noisy guidance and costly corrections. To address these issues, we propose a novel framework, bridging Vision and Text with LLMs for Few-Shot Learning (VT-FSL), which constructs precise cross-modal prompts conditioned on Large Language Models (LLMs) and support images, seamlessly integrating them through a geometry-aware alignment mechanism. It mainly consists of Cross-modal Iterative Prompting (CIP) and Cross-modal Geometric Alignment (CGA). Specifically, the CIP conditions an LLM on both class names and support images to generate precise class descriptions iteratively in a single structured reasoning pass. These descriptions not only enrich the semantic understanding of novel classes but also enable the zero-shot synthesis of semantically consistent images. The descriptions and synthetic images act respectively as complementary textual and visual prompts, providing high-level class semantics and low-level intra-class diversity to compensate for limited support data. Furthermore, the CGA jointly aligns the fused textual, support, and synthetic visual representations by minimizing the kernelized volume of the 3-dimensional parallelotope they span. It captures global and nonlinear relationships among all representations, enabling structured and consistent multimodal integration. The proposed VT-FSL method establishes new state-of-the-art performance across ten diverse benchmarks, including standard, cross-domain, and fine-grained few-shot learning scenarios. Code is available at https://github.com/peacelwh/VT-FSL.

We study how training data contributes to the emergence of toxic behaviors in large language models. Most prior work on reducing model toxicity adopts *reactive* approaches, such as fine-tuning pre-trained (and potentially toxic) models to align them with human values. In contrast, we propose a *proactive* approach—IF-Guide—that leverages influence functions to identify and suppress harmful tokens in the training data. To this end, we first show that standard influence functions are ineffective at discovering harmful training records. We then present a novel adaptation that measures token-level attributions from training data to model toxicity, along with techniques for selecting toxic training documents and a learning objective that can be integrated into both pre-training and fine-tuning. Moreover, IF-Guide does not rely on human-preference data, which is typically required by existing alignment methods. In our evaluation, we demonstrate that IF-Guide substantially reduces both explicit and implicit toxicity—by up to 10$\times$ compared to uncensored models, and up to 3$\times$ compared to baseline alignment methods such as DPO and RAD—across both pre-training and fine-tuning scenarios. IF-Guide is computationally efficient: a billion-parameter model is *not necessary* for computing influence scores; a million-parameter model—with 7.5$\times$ fewer parameters—can effectively serve as a proxy for identifying harmful data.

Uniform downsampling remains the de facto standard for reducing spatial resolution in vision backbones. In this work, we propose an alternative design built around a content-aware spatial grouping layer that dynamically assigns tokens to a reduced set based on image boundaries and their semantic content. Stacking our grouping layer across consecutive backbone stages results in hierarchical segmentation that arises *natively* in the feature extraction process, resulting in our coined Native Segmentation Vision Transformer. We show that a careful design of our architecture enables the emergence of strong segmentation masks solely from grouping layers, that is, without additional segmentation-specific heads. This sets the foundation for a new paradigm of *native*, backbone-level segmentation, which enables strong zero-shot results without mask supervision, as well as a minimal and efficient standalone model design for downstream segmentation tasks.

Supervised learning relies on high-quality labeled data, but obtaining such data through human annotation is both expensive and time-consuming. Recent work explores using large language models (LLMs) for annotation, but LLM-generated labels still fall short of human-level quality. To address this problem, we propose the Annotation with Critical Thinking (ACT) data pipeline, where LLMs serve not only as annotators but also as judges to critically identify potential errors. Human effort is then directed towards reviewing only the most "suspicious" cases, significantly improving the human annotation efficiency. Our major contributions are as follows: (1) ACT is applicable to a wide range of domains, including natural language processing (NLP), computer vision (CV), and multimodal understanding, by leveraging multimodal-LLMs (MLLMs). (2) Through empirical studies, we derive 7 insights on how to enhance annotation quality while efficiently reducing the human cost, and then translate these findings into user-friendly guidelines. (3) We theoretically analyze how to modify the loss function so that models trained on ACT data achieve similar performance to those trained on fully human-annotated data. Our experiments show that the performance gap can be reduced to less than 2% on most benchmark datasets while saving up to 90% of human costs.

Strategic classification (SC) explores how individuals or entities modify their features strategically to achieve favorable classification outcomes. However, existing SC methods, which are largely based on linear models or shallow neural networks, face significant limitations in terms of scalability and capacity when applied to real-world datasets with significantly increasing scale, especially in financial services and the internet sector. In this paper, we investigate how to leverage large language models to design a more scalable and efficient SC framework, especially in the case of growing individuals engaged with decision-making processes. Specifically, we introduce GLIM, a gradient-free SC method grounded in in-context learning. During the feed-forward process of self-attention, GLIM implicitly simulates the typical bi-level optimization process of SC, including both the feature manipulation and decision rule optimization. Without fine-tuning the LLMs, our proposed GLIM enjoys the advantage of cost-effective adaptation in dynamic strategic environments. Theoretically, we prove GLIM can support pre-trained LLMs to adapt to a broad range of strategic manipulations. We validate our approach through experiments with a collection of pre-trained LLMs on real-world and synthetic datasets in financial and internet domains, demonstrating that our GLIM exhibits both robustness and efficiency, and offering an effective solution for large-scale SC tasks.

Spatial Planning is a crucial part in the field of spatial intelligence, which requires the understanding and planning about object arrangements in space perspective. AI agents with the spatial planning ability can better adapt to various real-world applications, including robotic manipulation, automatic assembly, urban planning etc. Recent works have attempted to construct benchmarks for evaluating the spatial intelligence of Multimodal Large Language Models (MLLMs). Nevertheless, these benchmarks primarily focus on spatial reasoning based on typical Visual Question-Answering (VQA) forms, which suffers from the gap between abstract spatial understanding and concrete task execution. In this work, we take a step further to build a comprehensive benchmark called MineAnyBuild, aiming to evaluate the spatial planning ability of open-world AI agents in the Minecraft game. Specifically, MineAnyBuild requires an agent to generate executable architecture building plans based on the given multi-modal human instructions. It involves 4,000 curated spatial planning tasks and also provides a paradigm for infinitely expandable data collection by utilizing rich player-generated content. MineAnyBuild evaluates spatial planning through four core supporting dimensions: spatial understanding, spatial reasoning, creativity, and spatial commonsense. Based on MineAnyBuild, we perform a comprehensive evaluation for existing MLLM-based agents, revealing the severe limitations but enormous potential in their spatial planning abilities. We believe our MineAnyBuild will open new avenues for the evaluation of spatial intelligence and help promote further development for open-world AI agents capable of spatial planning.

Multimodal federated learning (MMFL) has gained increasing popularity due to its ability to leverage the correlation between various modalities, meanwhile preserving data privacy for different clients. However, recent studies show that correlation between modalities increase the vulnerability of federated learning against Gradient Inversion Attack (GIA). The complicated situation of MMFL privacy preserving can be summarized as follows: 1) different modality transmits different amounts of information, thus requires various protection strength; 2) correlation between modalities should be taken into account. This paper introduces an information theory perspective to analyze the leaked privacy in process of MMFL, and tries to propose a more reasonable protection method \textbf{Sec-MMFL} based on assessing different information leakage possibilities of each modality by conditional mutual information and adjust the corresponding protection strength. Moreover, we use mutual information to reduce the cross-modality information leakage in MMFL. Experiments have proven that our method can bring more balanced and comprehensive protection at an acceptable cost.

In classical AI, perception relies on learning state-based representations, while planning --- temporal reasoning over action sequences --- is typically achieved through search. We study whether such reasoning can instead emerge from representations that capture both perceptual and temporal structure. We show that standard temporal contrastive learning, despite its popularity, often fails to capture temporal structure due to its reliance on spurious features. To address this, we introduce Contrastive Representations for Temporal Reasoning (CRTR), a method that uses a negative sampling scheme to provably remove these spurious features and facilitate temporal reasoning. CRTR achieves strong results on domains with complex temporal structure, such as Sokoban and Rubik’s Cube. In particular, for the Rubik’s Cube, CRTR learns representations that generalize across all initial states and allow it to solve the puzzle using fewer search steps than BestFS — though with longer solutions. To our knowledge, this is the first method that efficiently solves arbitrary Cube states using only learned representations, without relying on an external search algorithm.

Modern AI systems, especially those interacting with the physical world, increasingly require real-time performance. However, the high latency of state-of-the-art generalist models, including recent vision-language-action models (VLAs), poses a significant challenge. While action chunking has enabled temporal consistency in high-frequency control tasks, it does not fully address the latency problem, leading to pauses or out-of-distribution jerky movements at chunk boundaries. This paper presents a novel inference-time algorithm that enables smooth asynchronous execution of action chunking policies. Our method, real-time chunking (RTC), is applicable to any diffusion- or flow-based VLA out of the box with no retraining. It generates the next action chunk while executing the current one, "freezing" actions guaranteed to execute and "inpainting" the rest. To test RTC, we introduce a new benchmark of 12 highly dynamic tasks in the Kinetix simulator, as well as evaluate 6 challenging real-world bimanual manipulation tasks. Results demonstrate that RTC is fast, performant, and uniquely robust to inference delay, significantly improving task throughput and enabling success in precise tasks --- such as lighting a match --- even in the presence of extreme latency.

Image restoration aims to recover sharp, high-quality images from degraded, low-quality inputs. Existing methods have progressively advanced from task-specific designs to general architectures, all-in-one frameworks, and composite degradation handling. Despite these advances, computational efficiency remains a critical factor for practical deployment. In this work, we present BioIR, an efficient and universal image restoration framework inspired by the human visual system. Specifically, we design two bio-inspired modules, Peripheral-to-Foveal (P2F) and Foveal-to-Peripheral (F2P), to emulate the perceptual processes of human vision, with a particular focus on the functional interplay between foveal and peripheral pathways. P2F delivers large-field contextual signals to foveal regions based on pixel-to-region affinity, while F2P propagates fine-grained spatial details through a static-to-dynamic two-stage integration strategy. Leveraging the biologically motivated design, BioIR achieves state-of-the-art performance across three representative image restoration settings: single-degradation, all-in-one, and composite degradation. Moreover, BioIR maintains high computational efficiency and fast inference speed, making it highly suitable for real-world applications.

We propose the Multi-modal Untrimmed Video Retrieval task, along with a new benchmark (MUVR) to advance video retrieval for long-video platforms. MUVR aims to retrieve untrimmed videos containing relevant segments using multi-modal queries. It has the following features: **1) Practical retrieval paradigm:** MUVR supports video-centric multi-modal queries, expressing fine-grained retrieval needs through long text descriptions, video tag prompts, and mask prompts. It adopts a one-to-many retrieval paradigm and focuses on untrimmed videos, tailored for long-video platform applications. **2) Multi-level visual correspondence:** To cover common video categories (e.g., news, travel, dance) and precisely define retrieval matching criteria, we construct multi-level visual correspondence based on core video content (e.g., news events, travel locations, dance moves) which users are interested in and want to retrieve. It covers six levels: copy, event, scene, instance, action, and others. **3) Comprehensive evaluation criteria:** We develop 3 versions of MUVR (i.e., Base, Filter, QA). MUVR-Base/Filter evaluates retrieval models, while MUVR-QA assesses MLLMs in a question-answering format. We also propose a Reranking Score to evaluate the reranking ability of MLLMs. MUVR consists of 53K untrimmed videos from the video platform Bilibili, with 1,050 multi-modal queries and 84K matches. Extensive evaluations of 3 state-of-the-art video retrieval models, 6 image-based VLMs, and 10 MLLMs are conducted. MUVR reveals the limitations of retrieval methods in processing untrimmed videos and multi-modal queries, as well as MLLMs in multi-video understanding and reranking. Our code and benchmark is available at https://github.com/debby-0527/MUVR.

The advent of Multimodal Large Language Models (MLLMs) has expanded AI capabilities to visual modalities, yet existing evaluation benchmarks remain limited to single-video understanding, overlooking the critical need for multi-video understanding in real-world scenarios (e.g., sports analytics and autonomous driving). To address this significant gap, we introduce **MVU-Eval**, the first comprehensive benchmark for evaluating **M**ulti-**V**ideo **U**nderstanding for MLLMs. Specifically, our MVU-Eval mainly assesses eight core competencies through 1,824 meticulously curated question-answer pairs spanning 4,959 videos from diverse domains, addressing both fundamental perception tasks and high-order reasoning tasks. These capabilities are rigorously aligned with real-world applications such as multi-sensor synthesis in autonomous systems and cross-angle sports analytics. Through extensive evaluation of state-of-the-art open-source and closed-source models, we reveal significant performance discrepancies and limitations in current MLLMs' ability to perform understanding across multiple videos. The benchmark will be made publicly available to foster future research.

We propose BlockScan, a customized Transformer for anomaly detection in blockchain transactions. Unlike existing methods that rely on rule-based systems or directly apply off-the-shelf large language models (LLMs), BlockScan introduces a series of customized designs to effectively model the unique data structure of blockchain transactions. First, a blockchain transaction is multi-modal, containing blockchain-specific tokens, texts, and numbers. We design a novel modularized tokenizer to handle these multi-modal inputs, balancing the information across different modalities. Second, we design a customized masked language modeling mechanism for pretraining the Transformer architecture, incorporating RoPE embedding and FlashAttention for handling longer sequences. Finally, we design a novel anomaly detection method based on the model outputs. We further provide theoretical analysis for the detection method of our system. Extensive evaluations on Ethereum and Solana transactions demonstrate BlockScan's exceptional capability in anomaly detection while maintaining a low false positive rate. Remarkably, BlockScan is the only method that successfully detects anomalous transactions on Solana with high accuracy, whereas all other approaches achieved very low or zero detection recall scores. This work sets a new benchmark for applying Transformer-based approaches in blockchain data analysis.