NeurIPS 2025 Papers

Video Temporal Grounding (TG) aims to temporally locate video segments matching a natural language description (a query) in a long video. While Vision-Language Models (VLMs) are effective at holistic semantic matching, they often struggle with fine-grained temporal localisation. Recently, Group Relative Policy Optimisation (GRPO) reformulates the inference process as a reinforcement learning task, enabling fine-grained grounding and achieving strong in-domain performance. However, GRPO relies on labelled data, making it unsuitable in unlabelled domains. Moreover, because videos are large and expensive to store and process, performing full-scale adaptation introduces prohibitive latency and computational overhead, making it impractical for real-time deployment. To overcome both problems, we introduce a Data-Efficient Unlabelled Cross-domain Temporal Grounding method, from which a model is first trained on a labelled source domain, then adapted to a target domain using only a small number of {\em unlabelled videos from the target domain}. This approach eliminates the need for target annotation and keeps both computational and storage overhead low enough to run in real time. Specifically, we introduce \textbf{U}ncertainty-quantified \textbf{R}ollout \textbf{P}olicy \textbf{A}daptation (\textbf{URPA}) for cross-domain knowledge transfer in learning video temporal grounding without target labels. URPA generates multiple candidate predictions using GRPO rollouts, averages them to form a pseudo label, and estimates confidence from the variance across these rollouts. This confidence then weights the training rewards, guiding the model to focus on reliable supervision. Experiments on three datasets across six cross-domain settings show that URPA generalises well using only a few unlabelled target videos. Codes are given in supplemental materials.

Learning rate is widely regarded as crucial for effective foundation model pretraining. Recent research explores and demonstrates the transferability of learning rate configurations across varying model and dataset sizes, etc. Nevertheless, these approaches are constrained to specific training scenarios and typically necessitate extensive hyperparameter tuning on proxy models. In this work, we propose \textbf{AdaLRS}, a plug-in-and-play adaptive learning rate search algorithm that conducts online optimal learning rate search via optimizing loss descent velocities. We provide theoretical and experimental analyzes to show that foundation model pretraining loss and its descent velocity are both convex and share the same optimal learning rate. Relying solely on training loss dynamics, AdaLRS involves few extra computations to guide the search process, and its convergence is guaranteed via theoretical analysis. Experiments on both LLM and VLM pretraining show that AdaLRS adjusts suboptimal learning rates to the neighborhood of optimum with marked efficiency and effectiveness, with model performance improved accordingly. We also show the robust generalizability of AdaLRS across varying training scenarios, such as different model sizes, training paradigms, base learning rate scheduler choices, and hyperparameter settings.

Multimodal large language models (MLLMs) face an inherent trade-off between faithfulness and creativity, as different tasks require varying degrees of associative reasoning. However, existing methods lack the flexibility to modulate this reasoning strength, limiting MLLMs' adaptability across factual and creative scenarios. To bridge this gap, we propose equipping MLLMs with mechanisms that enable flexible control over associative reasoning. We begin by investigating the internal mechanisms underlying associative behavior in MLLMs and find that: (1) middle layers play a pivotal role in shaping model’s associative tendencies, (2) modifying representations in these layers effectively regulates associative reasoning strength, and (3) hallucinations can be exploited to derive steering vectors that guide this modulation. Building on these findings, we introduce Flexible Association Control (FlexAC), a lightweight and training-free framework for modulating associative behavior in MLLMs. FlexAC first induces hallucination-guided intermediate representations to encode associative directions. Then, it selects high-association instances to construct effective associative steering vectors, whose strengths are adaptively calibrated to balance creative guidance with output stability. Finally, recognizing the multi-dimensional nature of associative reasoning, FlexAC incorporates task-specific associative vectors derived from a forward pass on a few target-domain samples, enabling models to follow diverse associative directions and better adapt to creative tasks. Notably, our method achieves up to a 5.8× improvement in creativity on Creation-MMBench and a 29\% reduction in hallucination rate on CHAIR, surpassing existing baselines and demonstrating its effectiveness in enabling flexible control over associative reasoning in MLLMs. Our code is available at https://github.com/ylhz/FlexAC.

Recent studies have demonstrated the efficacy of using Reinforcement Learning (RL) in building reasoning models that articulate chains of thoughts prior to producing final answers. However, despite ongoing advances that aim at enabling reasoning for vision-language tasks, existing open-source visual reasoning models typically generate reasoning content with pure natural language, lacking explicit integration of visual information. This limits their ability to produce clearly articulated and visually grounded reasoning chains. To this end, we propose Grounded Reasoning with Images and Texts (GRIT), a novel method for training MLLMs to think with images. GRIT introduces a grounded reasoning paradigm, in which models generate reasoning chains that interleave natural language and explicit bounding box coordinates. These coordinates point to regions of the input image that the model consults during its reasoning process. Additionally, GRIT is equipped with a reinforcement learning approach, GRPO-GR, built upon the GRPO algorithm. GRPO-GR employs robust rewards focused on the final answer accuracy and format of the grounded reasoning output, which eliminates the need for data with reasoning chain annotations or explicit bounding box labels. As a result, GRIT achieves exceptional data efficiency, requiring as few as 20 image-question-answer triplets from existing datasets. Comprehensive evaluations demonstrate that GRIT effectively trains MLLMs to produce coherent and visually grounded reasoning chains, showing a successful unification of reasoning and grounding abilities. All code, data, and checkpoints will be released.

The rapid progress of large language models (LLMs) has catalyzed the emergence of multimodal large language models (MLLMs) that unify visual understanding and image generation within a single framework. However, most existing MLLMs rely on autoregressive (AR) architectures, which impose inherent limitations on future development, such as the raster-scan order in image generation and restricted reasoning abilities in causal context modeling. In this work, we challenge the dominance of AR-based approaches by introducing FUDOKI, a unified multimodal model purely based on discrete flow matching, as an alternative to conventional AR paradigms. By leveraging metric-induced probability paths with kinetic optimal velocities, our framework goes beyond the previous masking-based corruption process, enabling iterative refinement with self-correction capability and richer bidirectional context integration during generation. To mitigate the high cost of training from scratch, we initialize FUDOKI from pre-trained AR-based MLLMs and adaptively transition to the discrete flow matching paradigm. Experimental results show that FUDOKI achieves performance comparable to state-of-the-art AR-based MLLMs across both visual understanding and image generation tasks, highlighting its potential as a foundation for next-generation unified multimodal models. Furthermore, we show that applying test-time scaling techniques to FUDOKI yields significant performance gains, further underscoring its promise for future enhancement through reinforcement learning.

Multiple Instance Learning (MIL) is a standard weakly supervised approach for Whole Slide Image (WSI) classification, where performance hinges on both feature representation and MIL pooling strategies. Recent research has predominantly focused on Transformer-based architectures adapted for WSIs. However, we argue that this trend faces a fundamental limitation: data scarcity. In typical settings, Transformer models yield only marginal gains without access to large-scale datasets—resources that are virtually inaccessible to all but a few well-funded research labs. Motivated by this, we revisit simple, non-attention MIL with unsupervised slide features and analyze temperature-$\beta$-controlled log-sum-exp (LSE) pooling. For slides partitioned into $N$ patches, we theoretically show that LSE has a smooth transition at a critical $\beta_{\mathrm{crit}}=\mathcal{O}(\log N)$ threshold, interpolating between mean-like aggregation (stable, better generalization but less sensitive) and max-like aggregation (more sensitive but looser generalization bounds). Grounded in this analysis, we introduce Maxsoft—a novel MIL pooling function that enables flexible control over this trade-off, allowing adaptation to specific tasks and datasets. To further tackle real-world deployment challenges such as specimen heterogeneity, we propose PerPatch augmentation—a simple yet effective technique that enhances model robustness. Empirically, Maxsoft achieves state-of-the-art performance in low-data regimes across four major benchmarks (CAMELYON16, CAMELYON17, TCGA-Lung, and SICAP-MIL), often matching or surpassing large-scale foundation models. When combined with PerPatch augmentation, this performance is further improved through increased robustness. Code is available at \href{https://github.com/jafarinia/maxsoft}{\texttt{https://github.com/jafarinia/maxsoft}}

We present CrypticBio, the largest publicly available multimodal dataset of visually confusing species, specifically curated to support the development of AI models in the context of biodiversity applications. Visually confusing or cryptic species are groups of two or more taxa that are nearly indistinguishable based on visual characteristics alone. While much existing work addresses taxonomic identification in a broad sense, datasets that directly address the morphological confusion of cryptic species are small, manually curated, and target only a single taxon. Thus, the challenge of identifying such subtle differences in a wide range of taxa remains unaddressed. Curated from real-world trends in species misidentification among community annotators of iNaturalist, CrypticBio contains 52K unique cryptic groups spanning 67K species represented in 166 million images. Records in the dataset include research-grade image annotations—scientific, multicultural, and multilingual species terminology, hierarchical taxonomy, spatiotemporal context, and associated cryptic groups. To facilitate easy subset curation from CrypticBio, we provide an open-source pipeline, CrypticBio-Curate. The multimodal design of the dataset provides complementary cues such as spatiotemporal context that support the identification of cryptic species. To highlight the importance of the dataset, we benchmark a suite of state-of-the-art foundation models across CrypticBio subsets of common, unseen, endangered, and invasive species, and demonstrate the substantial impact of spatiotemporal context on vision-language zero-shot learning for cryptic species. By introducing CrypticBio, we aim to catalyze progress toward real-world-ready fine-grained species classification models for biodiversity monitoring capable of handling the nuanced challenges of species ambiguity. The data and the code are publicly available in the project website https://georgianagmanolache.github.io/crypticbio.

There has recently been a surge of interest in Wearable Assistant Agents: agents embodied in a wearable form factor such as smart glasses, who can take actions toward a user’s stated goal — a high-level language-expressed command such as “where did I leave my keys?”, “Text Alice I will be late”, or “What’s the weather in Cancun?”. In this work, we consider the complementary problem of eliminating the effort required to interact with such an agent by proactively inferring the user’s goal from multimodal contextual observations. As vision-language models (VLMs) hold strong potential to ultimately solve this problem, our work focuses on creating a strong benchmark to measure progress toward this end. Given the limited prior work in this area, establishing the benchmark required collecting a novel multimodal goal-inference dataset; our dataset comprises ~30 hours of data from 363 participants across 3,482 recordings, featuring ground-truth reference goals alongside accompanying visual, audio, digital, and longitudinal contextual observations. We ran a human predictability study, where we found that humans set a strong baseline that comprises a de facto upper bound on model performance: they show multiple choice question (MCQ) accuracy of 93%, with the best VLM achieving about 84% accuracy. However, MCQ assesses discrimination, not the model’s ultimate task of generating the goal through open-ended text generation. Through a meta-evaluation, we find that a VLM judging the generated goals is as good as a human judge if it has access to a human-authored script of the video or a correct reference goal. Finally, we evaluate several families of modern vision-language models on the benchmark, showing that larger models have a significant performance advantage, but are still far from being practically useful, as they produce relevant goals only ~57% of the time. The best-performing smaller models—whose size makes them better suited to wearable applications—perform significantly worse than their counterparts, generating ~49% accuracy on the benchmark. Through a modality ablation, we show that models benefit from extra information in relevant modalities with minimal performance degradation from irrelevant modalities, but don’t gain as much when noisy modalities are included (e.g., in the case of digital context when most of the app state is irrelevant).

Fine-grained open-vocabulary object detection (FG-OVD) aims to detect novel object categories described by attribute-rich texts. While existing open-vocabulary detectors show promise at the base-category level, they underperform in fine-grained settings due to the semantic entanglement of subjects and attributes in pretrained vision-language model (VLM) embeddings -- leading to over-representation of attributes, mislocalization, and semantic drift in embedding space. We propose GUIDED, a decomposition framework specifically designed to address the semantic entanglement between subjects and attributes in fine-grained prompts. By separating object localization and fine-grained recognition into distinct pathways, GUIDED aligns each subtask with the module best suited for its respective roles. Specifically, given a fine-grained class name, we first use a language model to extract a coarse-grained subject and its descriptive attributes. Then the detector is guided solely by the subject embedding, ensuring stable localization unaffected by irrelevant or overrepresented attributes. To selectively retain helpful attributes, we introduce an attribute embedding fusion module that incorporates attribute information into detection queries in an attention-based manner. This mitigates over-representation while preserving discriminative power. Finally, a region-level attribute discrimination module compares each detected region against full fine-grained class names using a refined vision-language model with a projection head for improved alignment. Extensive experiments on FG-OVD and 3F-OVD benchmarks show that GUIDED achieves new state-of-the-art results, demonstrating the benefits of disentangled modeling and modular optimization.

Decision-making models for individuals, particularly in high-stakes scenarios like vaccine uptake, often diverge from population optimal predictions. This gap arises from the uniqueness of the individual decision-making process, shaped by numerical attributes (e.g., cost, time) and linguistic influences (e.g., personal preferences and constraints). Developing upon Utility Theory and leveraging the textual-reasoning capabilities of Large Language Models (LLMs), this paper proposes an Adaptive Textual-symbolic Human-centric Reasoning framework (ATHENA) to address the optimal information integration. ATHENA uniquely integrates two stages: First, it discovers robust, group-level symbolic utility functions via LLM-augmented symbolic discovery; Second, it implements individual-level semantic adaptation, creating personalized semantic templates guided by the optimal utility to model personalized choices. Validated on real-world travel mode and vaccine choice tasks, ATHENA consistently outperforms utility-based, machine learning, and other LLM-based models, lifting F1 score by at least 6.5\% over the strongest cutting-edge models. Further, ablation studies confirm that both stages of ATHENA are critical and complementary, as removing either clearly degrades overall predictive performance. By organically integrating symbolic utility modeling and semantic adaptation, ATHENA provides a new scheme for modeling human-centric decisions. The project page can be found at https://yibozh.github.io/Athena.

Chest X-rays (CXRs) are the most frequently performed imaging examinations in clinical settings. Recent advancements in Medical Large Multimodal Models (MLMMs) have enabled automated CXR interpretation, improving diagnostic accuracy and efficiency. However, despite their strong visual understanding, current MLMMs still face two major challenges: (1) insufficient region-level understanding and interaction, and (2) limited accuracy and interpretability due to single-step prediction. In this paper, we address these challenges by empowering MLMMs with anatomy-centric reasoning capabilities to enhance their interactivity and explainability. Specifically, we propose an Anatomical Ontology-Guided Reasoning (AOR) framework that accommodates both textual and optional visual prompts, centered on region-level information to enable multimodal multi-step reasoning. We also develop AOR-Instruction, a large instruction dataset for MLMMs training, under the guidance of expert physicians. Our experiments demonstrate AOR's superior performance in both Visual Question Answering (VQA) and report generation tasks. Code and data are available at: https://github.com/Liqq1/AOR.

Recent advancements in vision-language models (VLMs) have improved performance by increasing the number of visual tokens, which are often significantly longer than text tokens. However, we observe that most real-world scenarios do not require such an extensive number of visual tokens. While the performance drops significantly in a small subset of OCR-related tasks, models still perform accurately in most other general VQA tasks with only 1/4 resolution. Therefore, we propose to dynamically process distinct samples with different resolutions, and present a new paradigm for visual token compression, namely, VisionThink. It starts with a downsampled image and smartly decides whether it is sufficient for problem solving. Otherwise, the model could output a special token to request the higher-resolution image. Compared to existing Efficient VLM methods that compress tokens using fixed pruning ratios or thresholds, VisionThink autonomously decides whether to compress tokens case by case. As a result, it demonstrates strong fine-grained visual understanding capability on OCR-related tasks, and meanwhile saves substantial visual tokens on simpler tasks. We adopt reinforcement learning and propose the LLM-as-Judge strategy to successfully apply RL to general VQA tasks. Moreoever, we carefully design a reward function and penalty mechanism to achieve a stable and reasonable image resize call ratio. Extensive experiments demonstrate the superiority, efficiency, and effectiveness of our method. All our code and data are open-sourced.

Survey papers are foundational to the scholarly progress of research communities, offering structured overviews that guide both novices and experts across disciplines. However, the recent surge of AI-generated surveys, especially enabled by large language models (LLMs), has transformed this traditionally labor-intensive genre into a low-effort, high-volume output. While such automation lowers entry barriers, it also introduces a critical threat: the phenomenon we term the "survey paper DDoS attack" to the research community. This refers to the unchecked proliferation of superficially comprehensive but often redundant, low-quality, or even hallucinated survey manuscripts, which floods preprint platforms, overwhelms researchers, and erodes trust in the scientific record. In this position paper, we argue that we must stop uploading massive amounts of AI-generated survey papers (i.e., survey paper DDoS attack) to the research community, by instituting strong norms for AI-assisted review writing. We call for restoring expert oversight and transparency in AI usage and, moreover, developing new infrastructures such as Dynamic Live Surveys, community-maintained, version-controlled repositories that blend automated updates with human curation. Through quantitative trend analysis, quality audits, and cultural impact discussion, we show that safeguarding the integrity of surveys is no longer optional but imperative to the research community.

Recent advances in self-supervised learning for Vision Transformers (ViTs) have fueled breakthroughs in remote sensing (RS) foundation models. However, the quadratic complexity of self-attention poses a significant barrier to scalability, particularly for large models and high-resolution images. While the linear-complexity Mamba architecture offers a promising alternative, existing RS applications of Mamba remain limited to supervised tasks on small, domain-specific datasets. To address these challenges, we propose RoMA, a framework that enables scalable self-supervised pretraining of Mamba-based RS foundation models using large-scale, diverse, unlabeled data. RoMA enhances scalability for high-resolution images through a tailored auto-regressive learning strategy, incorporating two key innovations: 1) a rotation-aware pretraining mechanism combining adaptive cropping with angular embeddings to handle sparsely distributed objects with arbitrary orientations, and 2) multi-scale token prediction objectives that address the extreme variations in object scales inherent to RS imagery. Systematic empirical studies validate that Mamba adheres to RS data and parameter scaling laws, with performance scaling reliably as model and data size increase. Furthermore, experiments across scene classification, object detection, and semantic segmentation tasks demonstrate that RoMA-pretrained Mamba models consistently outperform ViT-based counterparts in both accuracy and computational efficiency. The source code and pretrained models have be released at https://github.com/MiliLab/RoMA.

In this work, we demonstrate that affine mappings between residual streams of language models is a cheap way to effectively transfer represented features between models. We apply this technique to transfer the \textit{weights} of Sparse Autoencoders (SAEs) between models of different sizes to compare their representations. We find that small and large models learn highly similar representation spaces, which motivates training expensive components like SAEs on a smaller model and transferring to a larger model at a FLOPs savings. For example, using a small-to-large transferred SAE as initialization can lead to 50% cheaper training runs when training SAEs on larger models. Next, we show that transferred probes and steering vectors can effectively recover ground truth performance. Finally, we dive deeper into feature-level transferability, finding that semantic and structural features transfer noticeably differently while specific classes of functional features have their roles faithfully mapped. Overall, our findings illustrate similarities and differences in the linear representation spaces of small and large models and demonstrate a method for improving the training efficiency of SAEs.

The dynamic nature of information necessitates continuously updating large vision-language models (LVLMs). While recent knowledge editing techniques hint at promising directions, they often focus on editing a single modality (vision or language) in isolation. This prevalent practice neglects the inherent multimodality of LVLMs and the continuous nature of knowledge updates, potentially leading to suboptimal editing outcomes when considering the interplay between modalities and the need for ongoing knowledge refinement. To address these limitations, we propose MemEIC, a novel method for Continual and Compositional Knowledge Editing (CCKE) in LVLMs. MemEIC enables compositional editing of both visual and textual knowledge sequentially. Our approach employs a hybrid external-internal editor featuring a dual external memory for cross-modal evidence retrieval and dual LoRA adapters that facilitate disentangled parameter updates for each modality. A key component is a brain-inspired knowledge connector, activated selectively for compositional reasoning, that integrates information across different modalities. Experiments demonstrate that MemEIC significantly improves performance on complex multimodal questions and effectively preserves prior edits, setting a new benchmark for CCKE in LVLMs. Our project is available at https://github.com/MemEIC/MemEIC.

Transferability estimation identifies the best pre-trained models for downstream tasks without incurring the high computational cost of full fine-tuning. This capability facilitates deployment and advances the pre-training and fine-tuning paradigm. However, existing methods often struggle to accurately assess transferability for emerging pre-trained models with diverse architectures, training strategies, and task alignments. In this work, we propose Implicit Transferability Modeling (ITM), a novel framework that implicitly models each model’s intrinsic transferability, coupled with a Divide-and-Conquer Variational Approximation (DVA) strategy to efficiently approximate embedding space evolution. This design enables generalization across a broader range of models and downstream tasks. Extensive experiments on a comprehensive benchmark—spanning extensive training regimes and a wider variety of model types—demonstrate that ITM consistently outperforms existing methods in terms of stability, effectiveness, and efficiency.

Most video-anomaly research stops at frame-wise detection, offering little insight into why an event is abnormal, typically outputting only frame-wise anomaly scores without spatial or semantic context. Recent video anomaly localization and video anomaly understanding methods improve explainability but remain data-dependent and task-specific. We propose a unified reasoning framework that bridges the gap between temporal detection, spatial localization, and textual explanation. Our approach is built upon a chained test-time reasoning process that sequentially connects these tasks, enabling holistic zero-shot anomaly analysis without any additional training. Specifically, our approach leverages intra-task reasoning to refine temporal detections and inter-task chaining for spatial and semantic understanding, yielding improved interpretability and generalization in a fully zero-shot manner. Without any additional data or gradients, our method achieves state-of-the-art zero-shot performance across multiple video anomaly detection, localization, and explanation benchmarks. The results demonstrate that careful prompt design with task-wise chaining can unlock the reasoning power of foundation models, enabling practical, interpretable video anomaly analysis in a fully zero-shot manner. Project Page: https://rathgrith.github.io/Unified_Frame_VAA/.

In this paper, we study the visual redundancy problem of multimodal large language models (MLLMs) from the perspective of attention behaviors. Via extensive empirical experiments, we observe and conclude three main inference stages of MLLMs: (i) Early fusion between tokens is first accomplished quickly. (ii) Intra-modality modeling then comes to play. (iii) Multimodal reasoning resumes and lasts until the end of inference. In particular, we reveal that visual tokens will stop contributing to reasoning when the text tokens receive enough image information. Based on this observation, we propose an effective method to improve the efficiency of MLLMs, termed dynamic visual-token exit (DyVTE), which is orthogonal but collaborative to previous token-wise visual compression methods. To validate the efficacy of DyVTE, we apply it to a set of MLLMs, including LLaVA, VILA, EAGLE and InternVL. The experimental results not only show the effectiveness of our DyVTE in improving MLLMs' efficiency, e.g., DyVTE reduces the computation overhead of LLaVA-1.5 by up to 45.7% without performance drop, but also reveal a general pattern across multiple MLLMs, well facilitating the in-depth analysis of MLLMs. Our code is anonymously released at https://anonymous.4open.science/r/AnonymousDyVTE-26AB/.