NeurIPS 2025 Papers

Modern AI systems struggle most in environments where reliability is critical - scenes with smoke, poor visibility, and structural deformation. Each year, tens of thousands of firefighters are injured on duty, often due to breakdowns in situational perception. We introduce Fire360, a benchmark for evaluating perception and reasoning in safety-critical firefighting scenarios. The dataset includes 228 360° videos from professional training sessions under diverse conditions (e.g., low light, thermal distortion), annotated with action segments, object locations, and degradation metadata. Fire360 supports five tasks: Visual Question Answering, Temporal Action Captioning, Object Localization, Safety-Critical Reasoning, and Transformed Object Retrieval (TOR). TOR tests whether models can match pristine exemplars to fire-damaged counterparts in unpaired scenes, evaluating episodic memory under irreversible visual transformations. While human experts achieve 83.5% on TOR, models like GPT-4o lag significantly, exposing failures in reasoning under degradation. By releasing Fire360 and its evaluation suite, we aim to advance models that not only see, but also remember, reason, and act under uncertainty. The dataset is available at https://uofi.box.com/v/fire360dataset

Multimodal learning systems often encounter challenges related to modality imbalance, where a dominant modality may overshadow others, thereby hindering the learning of weak modalities. Conventional approaches often force weak modalities to align with dominant ones in "Learning to be (the same)" (Positive Learning), which risks suppressing the unique information inherent in the weak modalities. To address this challenge, we offer a new learning paradigm: "Learning Not to be" (Negative Learning). Instead of enhancing weak modalities’ target-class predictions, the dominant modalities dynamically guide the weak modality to suppress non-target classes. This stabilizes the decision space and preserves modality-specific information, allowing weak modalities to preserve unique information without being over-aligned. We proceed to reveal the multimodal learning from a robustness perspective and theoretically derive the Multimodal Negative Learning (MNL) framework, which introduces a dynamic guidance mechanism tailored for negative learning. Our method provably tightens the robustness lower bound of multimodal learning by increasing the Unimodal Confidence Margin (UCoM) and reduces the empirical error of weak modalities, particularly under noisy and imbalanced scenarios. Extensive experiments across multiple benchmarks demonstrate the effectiveness and generalizability of our approach against the competing methods. The code will be available at: https://github.com/BaoquanGong/Multimodal-Negative-Learning.git

Quantized training of Large Language Models (LLMs) remains an open challenge, as maintaining accuracy while performing all matrix multiplications in low precision has proven difficult. This is particularly the case when fine-tuning pre-trained models, which can have large weight, activation, and error (output gradient) outlier values that make lower-precision optimization difficult. To address this, we present HALO, a new quantization-aware training approach for Transformers that enables accurate and efficient low-precision training by combining 1) strategic placement of Hadamard rotations in both forward and backward passes, which mitigate outliers, 2) high-performance kernel support, and 3) FSDP integration for low-precision communication. Our approach ensures that all large matrix multiplications during the forward and backward passes are executed in lower precision. Applied to LLaMa models, HALO achieves near-full-precision-equivalent results during fine-tuning on various tasks, while delivering up to 1.41x end-to-end speedup for full fine-tuning on RTX 4090 GPUs. HALO efficiently supports both standard and parameter-efficient fine-tuning (PEFT). Our results demonstrate the first practical approach to fully quantized LLM fine-tuning that maintains accuracy in INT8 and FP6 precision, while delivering performance benefits.

The Arrow of Time (AoT)—time's irreversible flow shaping physical events—is fundamental to video comprehension, yet remains a significant challenge for modern large multimodal models (LMMs). Current LMMs struggle to perceive and utilize temporal directionality in video when responding to language queries, obstructing deeper temporal understanding. We tackle this deficiency by first providing a critical analysis of existing benchmarks and models. We then introduce ArrowRL, a reinforcement learning (RL)-based training strategy with an innovative reverse reward that instills AoT awareness by encouraging divergent video interpretations between forward and reversed visual frames. For rigorous evaluation, we additionally develop AoTBench, a new multi-faceted benchmark probing temporally challenging questions. Experiments show ArrowRL greatly advances temporal perception: it not only achieves substantial improvements on our challenging AoTBench but also demonstrably boosts performance on standard video question answering (VQA) benchmarks (with peak accuracy gains reaching over 20% and 10% respectively). This validates ArrowRL's effectiveness and highlights the critical need for dedicated AoT understanding in LMMs.

Vision foundation models achieve remarkable performance but are only available in a limited set of pre-determined sizes, forcing sub-optimal deployment choices under real-world constraints. We introduce SnapViT: single-shot network approximation for pruned Vision Transformers, a new post-pretraining structured pruning method that enables elastic inference across a continuum of compute budgets. Our approach efficiently combines gradient information with cross-network structure correlations, approximated via an evolutionary algorithm, does not require labeled data, generalizes to models without a classification head, and is retraining-free. Experiments on DINO, SigLIPv2, DeIT, and AugReg models demonstrate superior performance over state-of-the-art methods across various sparsities, requiring less than five minutes on a single A100 GPU to generate elastic models that can be adjusted to any computational budget. Our key contributions include an efficient pruning strategy for pretrained Vision Transformers, a novel evolutionary approximation of Hessian off-diagonal structures, and a self-supervised importance scoring mechanism that maintains strong performance without requiring retraining or labels. Code and pruned models are available at: https://elastic.ashita.nl/

Vision-Language Foundation Models (VLMs), trained on large-scale multimodal datasets, have driven significant advances in Artificial Intelligence (AI) by enabling rich cross-modal reasoning. Despite their success in general domains, applying these models to medical imaging remains challenging due to the limited availability of diverse imaging modalities and multilingual clinical data. Most existing medical VLMs are trained on a subset of imaging modalities and focus primarily on high-resource languages, thus limiting their generalizability and clinical utility. To address these limitations, we introduce a novel Vietnamese-language multimodal medical dataset consisting of 2,757 whole-body PET/CT volumes from independent patients and their corresponding full-length clinical reports. This dataset is designed to fill two pressing gaps in medical AI development: (1) the lack of PET/CT imaging data in existing VLMs training corpora, which hinders the development of models capable of handling functional imaging tasks; and (2) the underrepresentation of low-resource languages, particularly the Vietnamese language, in medical vision-language research. To the best of our knowledge, this is the first dataset to provide comprehensive PET/CT-report pairs in Vietnamese. We further introduce a training framework to enhance VLMs' learning, including data augmentation and expert-validated test sets. We conduct comprehensive experiments benchmarking state-of-the-art VLMs on downstream tasks, including medical report generation and visual question answering. The experimental results show that incorporating our dataset significantly improves the performance of existing VLMs. However, despite these advancements, the models still underperform on clinically critical criteria, particularly the diagnosis of lung cancer, indicating substantial room for future improvement. We believe this dataset and benchmark will serve as a pivotal step in advancing the development of more robust VLMs for medical imaging, particularly in low-resource languages, and improving their clinical relevance in Vietnamese healthcare.

Time series foundation models (TSFMs) demonstrate impressive zero-shot performance for time series forecasting. However, an important yet underexplored challenge is how to effectively finetune TSFMs on specific downstream tasks. While naive finetuning can yield performance gains, we argue that it falls short of fully leveraging TSFMs' capabilities, often resulting in overfitting and suboptimal performance. Given the diverse temporal patterns across sampling scales and the inherent multi-scale forecasting capabilities of TSFMs, we adopt a causal perspective to analyze finetuning process, through which we highlight the critical importance of explicitly modeling multiple scales and reveal the shortcomings of naive approaches. Focusing on encoder-based TSFMs, we propose Multiscale finetuning (MSFT), a simple yet general framework that explicitly integrates multi-scale modeling into the finetuning process. Experimental results on three different backbones (Moirai, Moment and Units) demonstrate that TSFMs finetuned with MSFT not only outperform naive and typical parameter efficient finetuning methods but also surpass state-of-the-art deep learning methods. Codes are available at https://github.com/zqiao11/MSFT.

This position paper argues for two claims regarding AI testing and evaluation. First, to remain informative about deployment behaviour, evaluations need account for the possibility that AI systems understand their circumstances and reason strategically. Second, game-theoretic analysis can inform evaluation design by formalising and scrutinising the reasoning in evaluation-based safety cases. Drawing on examples from existing AI systems, a review of relevant research, and formal strategic analysis of a stylised evaluation scenario, we present evidence for these claims and motivate several research directions.

This study aims to enlarge our current knowledge on the application of brain-inspired network science principles for training artificial neural networks (ANNs) with sparse connectivity. Dynamic sparse training (DST) emulates the synaptic turnover of real brain networks, reducing the computational demands of training and inference in ANNs. However, existing DST methods face difficulties in maintaining peak performance at high connectivity sparsity levels. The Cannistraci-Hebb training (CHT) is a brain-inspired method that is used in DST for growing synaptic connectivity in sparse neural networks. CHT leverages a gradient-free, topology-driven link regrowth mechanism, which has been shown to achieve ultra-sparse (1\% connectivity or lower) advantage across various tasks compared to fully connected networks. Yet, CHT suffers two main drawbacks: (i) its time complexity is $\mathcal{O}(N\cdot d^3)$- N node network size, d node degree - hence it can be efficiently applied only to ultra-sparse networks. (ii) it rigidly selects top link prediction scores, which is inappropriate for the early training epochs, when the network topology presents many unreliable connections. Here, we design the first brain-inspired network model - termed bipartite receptive field (BRF) - to initialize the connectivity of sparse artificial neural networks. Then, we propose a matrix multiplication GPU-friendly approximation of the CH link predictor, which reduces the computational complexity to $\mathcal{O}(N^3)$, enabling a fast implementation of link prediction in large-scale models. Moreover, we introduce the Cannistraci-Hebb training soft rule (CHTs), which adopts a flexible strategy for sampling connections in both link removal and regrowth, balancing the exploration and exploitation of network topology. Additionally, we propose a sigmoid-based gradual density decay strategy, leading to an advanced framework referred to as CHTss. Empirical results show that BRF offers performance advantages over previous network science models. Using 1\% of connections, CHTs outperforms fully connected networks in MLP architectures on visual classification tasks, compressing some networks to less than 30\% of the nodes. Using 5\% of the connections, CHTss outperforms fully connected networks in two Transformer-based machine translation tasks. Finally, with only 30\% of the connections, both CHTs and CHTss achieve superior performance over other dynamic sparse training methods, and perform on par with—or even surpass—their fully connected counterparts in language modeling across various sparsity levels within the LLaMA model family. The code is available at: https://github.com/biomedical-cybernetics/Cannistraci-Hebb-training.

Weight-only quantization is widely used to mitigate the memory-bound nature of LLM inference. Codebook-based methods extend this trend by achieving strong accuracy in the extremely low-bit regime (e.g., 2-bit). However, current kernels rely on dequantization, which repeatedly fetches centroids and reconstructs weights, incurring substantial latency and cache pressure. We present CodeGEMM, a codebook-centric GEMM kernel that replaces dequantization with precomputed inner products between centroids and activations stored in a lightweight Psumbook. At inference, code indices directly gather these partial sums, eliminating per-element lookups and reducing the on-chip footprint. The kernel supports the systematic exploration of latency–memory–accuracy trade-offs under a unified implementation. On Llama-3 models, CodeGEMM delivers 1.83x (8B) and 8.93x (70B) speedups in the 2-bit configuration compared to state-of-the-art codebook-based quantization at comparable accuracy and further improves computing efficiency and memory subsystem utilization.

Spatial referring is a fundamental capability of embodied robots to interact with the 3D physical world. However, even with the powerful pretrained VLMs, recent approaches are still not qualified to accurately understand the complex 3D scenes and dynamically reason about the instruction-indicated locations for interaction. To this end, we propose RoboRefer, a 3D-aware vision language model (VLM) that can first achieve precise spatial understanding by integrating a disentangled but dedicated depth encoder via supervised fine-tuning (SFT). Moreover, RoboRefer advances generalized multi-step spatial reasoning via reinforcement fine-tuning (RFT), with metric-sensitive process reward functions tailored for spatial referring tasks. To support SFT and RFT training, we introduce RefSpatial, a large-scale dataset of 20M QA pairs (2x prior), covering 31 spatial relations (vs. 15 prior) and supporting complex reasoning processes (up to 5 steps). In addition, we introduce RefSpatial-Bench, a challenging benchmark filling the gap in evaluating spatial referring with multi-step reasoning. Experiments show that SFT-trained RoboRefer achieves state-of-the-art spatial understanding, with an average success rate of 89.6%. RFT-trained RoboRefer further outperforms all other baselines by a large margin, even surpassing Gemini-2.5-Pro by 12.4% in average accuracy on RefSpatial-Bench. Notably, RoboRefer can be integrated with various control policies to execute long-horizon, dynamic tasks across diverse robots (e,g., UR5, G1 humanoid) in cluttered real-world scenes.

Video Question Answering (VideoQA) task serves as a critical playground for evaluating whether foundation models can effectively perceive, understand, and reason about dynamic real-world scenarios. However, existing Multimodal Large Language Models (MLLMs) struggle with simultaneously ensuring the ability to model spatial relationships between video frames and to understand the causal dynamics of temporal evolution on complex and reasoning-intensive VideoQA. In this work, we equip MLLM with a comprehensive and extensible Video Toolkit, to enhance MLLM’s spatiotemporal reasoning capabilities as well as guarantee the harmony between the quantity and diversity of tools. To better control the tool invocation sequence and avoid toolchain shortcut issues, we propose a Spatiotemporal Reasoning Framework (STAR) that strategically schedules temporal and spatial tools, thereby progressively localizing the key area in the video. Our STAR framework enhances GPT-4o using lightweight tools, achieving an 8.2% gain on VideoMME and 4.6% on LongVideoBench. We believe that our proposed Video Toolkit and STAR framework make an important step towards building autonomous and intelligent video analysis assistants. The code is publicly available at https://github.com/fansunqi/VideoTool.

Test-time adaptation with pre-trained vision-language models shows impressive zero-shot classification abilities, and training-free methods further improve the performance without any optimization burden. However, existing training-free test-time adaptation methods typically rely on entropy criteria to select the visual features and update the visual caches, while ignoring the generalizable factors, such as shape-sensitive and style-insensitive factors. In this paper, we propose a novel shape and style guidance method (SSG) for training-free test-time adaptation in vision-language models, aiming to highlight the shape-sensitive (SHS) and style-insensitive (STI) factors in addition to entropy criteria. Specifically, SSG perturbs the raw test image with shape and style corruption operations, and measures the prediction difference between the raw and corrupted one as perturbed prediction difference (PPD). Based on the PPD measurement, SSG reweights the high-confidence visual features and corresponding predictions, aiming to highlight the effect of SHS and STI factors during the test-time procedure. Furthermore, SSG takes both PPD and entropy into consideration to update the visual cache, aiming to maintain the stored sample with high entropy and generalizable factors. Extensive experimental results on out-of-distribution and cross-domain benchmark datasets demonstrate that our proposed SSG consistently outperforms previous state-of-the-art methods while also exhibiting promising computational efficiency.

Large language model-assisted Evolutionary Search (LES) has emerged as a promising approach for Automated Algorithm Discovery (AAD). While many evolutionary search strategies have been developed for classic optimization problems, LES operates in abstract language spaces, presenting unique challenges for applying these strategies effectively. To address this, we propose a general LES framework that incorporates feature-assisted niche construction within abstract search spaces, enabling the seamless integration of niche-based search strategies from evolutionary computation. Building on this framework, we introduce PartEvo, an LES method that combines niche collaborative search and advanced prompting strategies to improve algorithm discovery efficiency. Experiments on both synthetic and real-world optimization problems show that PartEvo outperforms human-designed baselines and surpasses prior LES methods, such as Eoh and Funsearch. In particular, on resource scheduling tasks, PartEvo generates meta-heuristics with low design costs, achieving up to 90.1\% performance improvement over widely-used baseline algorithms, highlighting its potential for real-world applications.

Vision-Language Models (VLMs) acquire real-world knowledge and general reasoning ability through Internet-scale image-text corpora. They can augment robotic systems with scene understanding and task planning, and assist visuomotor policies that are trained on robot trajectory data. We explore the reverse paradigm — using rich, real, multi-modal robot trajectory data to enhance and evaluate VLMs. In this paper, we present Robo2VLM, a Visual Question Answering (VQA) dataset generation framework for VLMs. Given a human tele-operated robot demonstration with video and robot data, Robo2VLM derives ground-truth from non-visual and non-descriptive sensory modalities, such as end-effector pose, gripper aperture, and force sensing. Based on these modalities, it segments the robot trajectory into a sequence of manipulation phases. At each phase, Robo2VLM uses scene and interaction understanding to identify 3D properties of the robot, task goal, and the target object. The properties are used to generate representative VQA queries – images with textural multiple-choice questions – based on spatial, goal-conditioned, and interaction reasoning question templates. We use a subset of Open X-Embodiment to generate Robo2VLM-1, a large-scale in-the-wild dataset with 684,710 questions based on 463 distinct scenes and 3,396 robotic manipulation tasks from 176k real robot trajectories. Results suggest that Robo2VLM-1 can benchmark and improve VLM capabilities in spatial and interaction reasoning.

Alignment with human preferences is an important evaluation aspect of LLMs, requiring them to be helpful, honest, safe, and to precisely follow human instructions. Evaluating large language models' (LLMs) alignment typically involves directly assessing their open-ended responses, requiring human annotators or strong LLM judges. Conversely, LLMs themselves have also been extensively evaluated as judges for assessing alignment. In this work, we examine the relationship between LLMs' generation and evaluation capabilities in aligning with human preferences. To this end, we first conduct a comprehensive analysis of the generation-evaluation consistency (GE-consistency) among various LLMs, revealing a strong correlation between their generation and evaluation capabilities when evaluated by a strong LLM preference oracle (GPT-4o). Utilizing this finding, we propose a benchmarking paradigm that measures LLM alignment with human preferences without directly evaluating their generated outputs, instead assessing LLMs in their role as evaluators. Our evaluation shows that our proposed benchmark, AlignEval, matches or surpasses widely used automatic LLM evaluation benchmarks, such as AlpacaEval and Arena-Hard, in capturing human preferences when ranking LLMs. Our study offers valuable insights into the connection between LLMs' generation and evaluation capabilities, and introduces a benchmark that assesses alignment without directly evaluating model outputs.

Mamba, a lightweight sequence modeling framework offering near-linear complexity, presents a promising alternative to Transformers. In this work, we introduce MOGO (Mamba Only Glances Once), an end-to-end framework for efficient video action detection built entirely on the Mamba architecture. In MOGO, our newly designed Mamba-based decoder can even use just one Mamba layer to effectively perform action detection. It uses neither Transformer structures nor RCNN-like methods for proposal detection. Our framework introduces two key innovations. First, we propose a pure Mamba-based encoder-decoder architecture. The encoder processes cross-frame video information, while the decoder incorporates two novel Mamba-based structures that leverage Mamba’s intrinsic capabilities to detect actions. Theoretical analysis and ablation experiments confirm their synergy and the necessity of each structure. Second, we design a video token construction mechanism to improve the model's performance. The token importance block can ensure that the retained token information is highly relevant to the predicted targets. These two innovations make MOGO both efficient and accurate, as demonstrated on the JHMDB and UCF101-24 benchmark datasets. Compared to SOTA action detection methods, MOGO achieves superior performance in terms of GFLOPs, model parameters, and inference speed (latency) with comparable detection precision. Additionally, it requires significantly less GPU memory than some SOTA token reconstruction methods. Code is available at https://github.com/YunqingLiu-ML/MOGO.

Recent advancements in large language models have demonstrated how chain-of-thought (CoT) and reinforcement learning (RL) can improve performance. However, applying such reasoning strategies to the visual generation domain remains largely unexplored. In this paper, we present **T2I-R1**, a novel reasoning-enhanced text-to-image generation model, powered by RL with a bi-level CoT reasoning process. Specifically, we identify two levels of CoT that can be utilized to enhance different stages of generation: (1) the semantic-level CoT for high-level planning of the prompt and (2) the token-level CoT for low-level pixel processing during patch-by-patch generation. To better coordinate these two levels of CoT, we introduce **BiCoT-GRPO** with an ensemble of generation rewards, which seamlessly optimizes both generated CoTs within the same training step. By applying our reasoning strategies to the baseline model, Janus-Pro, we achieve superior performance with 13% improvement on T2I-CompBench and 19% improvement on the WISE benchmark, even surpassing the state-of-the-art model FLUX.1. All the training code is in the supplementary material and will be made public.

The Contrastive Language-Image Pretraining (CLIP) model has significantly advanced vision-language modeling by aligning image-text pairs from large-scale web data through self-supervised contrastive learning. Yet, its reliance on uncurated Internet-sourced data exposes it to data poisoning and backdoor risks. While existing studies primarily investigate image-based attacks, the text modality, which is equally central to CLIP's training, remains underexplored. In this work, we introduce ToxicTextCLIP, a framework for generating high-quality adversarial texts that target CLIP during the pre-training phase. The framework addresses two key challenges: semantic misalignment caused by background inconsistency with the target class, and the scarcity of background-consistent texts. To this end, ToxicTextCLIP iteratively applies: 1) a background-aware selector that prioritizes texts with background content aligned to the target class, and 2) a background-driven augmenter that generates semantically coherent and diverse poisoned samples. Extensive experiments on classification and retrieval tasks show that ToxicTextCLIP achieves up to 95.83\% poisoning success and 98.68% backdoor Hit@1, while bypassing RoCLIP, CleanCLIP and SafeCLIP defenses. The source code can be accessed via https://github.com/xinyaocse/ToxicTextCLIP/.