NeurIPS 2025 Papers

Despite promising performance on open-source large vision-language models (LVLMs), transfer-based targeted attacks often fail against closed-source commercial LVLMs. Analyzing failed adversarial perturbations reveals that the learned perturbations typically originate from a uniform distribution and lack clear semantic details, resulting in unintended responses. This critical absence of semantic information leads commercial black-box LVLMs to either ignore the perturbation entirely or misinterpret its embedded semantics, thereby causing the attack to fail. To overcome these issues, we propose to refine semantic clarity by encoding explicit semantic details within local regions, thus ensuring the capture of finer-grained features and inter-model transferability, and by concentrating modifications on semantically rich areas rather than applying them uniformly. To achieve this, we propose *a simple yet highly effective baseline*: at each optimization step, the adversarial image is cropped randomly by a controlled aspect ratio and scale, resized, and then aligned with the target image in the embedding space. While the naive source-target matching method has been utilized before in the literature, we are the first to provide a tight analysis, which establishes a close connection between perturbation optimization and semantics. Experimental results confirm our hypothesis. Our adversarial examples crafted with local-aggregated perturbations focused on crucial regions exhibit surprisingly good transferability to commercial LVLMs, including GPT-4.5, GPT-4o, Gemini-2.0-flash, Claude-3.5/3.7-sonnet, and even reasoning models like o1, Claude-3.7-thinking and Gemini-2.0-flash-thinking. Our approach achieves success rates exceeding 90\% on GPT-4.5, 4o, and o1, significantly outperforming all prior state-of-the-art attack methods with lower $\ell_1/\ell_2$ perturbations. Our optimized adversarial examples under different configurations are available at https://huggingface.co/datasets/MBZUAI-LLM/M-Attack_AdvSamples and our training code at https://github.com/VILA-Lab/M-Attack.

Foundation models have recently shown strong potential as web agents, capable of interpreting high-level instructions and interacting with complex web interfaces. However, existing training paradigms for these agents often rely on predefined task datasets and curated demonstrations, limiting their scalability, adaptability, and capacity for self-improvement. In this work, we introduce *Self-guided hierArchical exploration for Generalist wEb agents* (SAGE), a new training framework designed to support autonomous skill acquisition through self-guided hierarchical exploration. Our method introduces a three-tier exploration strategy: a pre-exploration phase to build structural understanding of web environments, a top-level exploration strategy to generate a self-evolving curriculum of tasks from easy to hard, and a low-level exploration mechanism that combines planning-based rollouts with step-wise learning to improve policy efficiency. Together, these components form a scalable, supervision-free framework for web agent training. Experimental results on WebVoyager and WebArena demonstrate that our method significantly outperforms prior approaches, enabling foundation model agents to learn complex web tasks with greater generalization and robustness.

Existing AutoML systems have advanced the automation of machine learning (ML); however, they still require substantial manual configuration and expert input, particularly when handling multimodal data. We introduce MLZero, a novel multi-agent framework powered by Large Language Models (LLMs) that enables end-to-end ML automation across diverse data modalities with minimal human intervention. A cognitive perception module is first employed, transforming raw multimodal inputs into perceptual context that effectively guides the subsequent workflow. To address key limitations of LLMs, such as hallucinated code generation and outdated API knowledge, we enhance the iterative code generation process with semantic and episodic memory. MLZero demonstrates superior performance on MLE-Bench Lite, outperforming all competitors in both success rate and solution quality, securing six gold medals. Furthermore, when evaluated on our Multimodal AutoML Agent Benchmark, which includes 25 more challenging tasks spanning diverse data modalities, MLZero outperforms the competing methods by a large margin with a success rate of 0.92 (+263.6\%) and an average rank of 2.28. Our approach maintains its robust effectiveness even with a compact 8B LLM, outperforming full-size systems from existing solutions.

Inductive reasoning infers general rules from observed evidence, which is one of the most critical intelligence abilities. Previous works have succeeded in formal languages but suffer from onerous and error-prone conversions between a particular formal language and the working language. As large language models (LLMs) have emerged, direct reasoning with various kinds of languages, especially natural languages, without formal language involvement has become feasible. However, existing LLM-based inductive reasoning usually relies on LLM's intrinsic generation ability, which is prone to LLM's hallucination and lacks systematic guidance according to the nature of inductive reasoning. To this end, we propose HypoBootstrap, an integrated framework for inductive reasoning that generates and confirms hypotheses both in a bootstrapping manner. Regarding hypothesis generation, we propose a novel bootstrapping generation strategy, bootstrapping object hypotheses, relational hypotheses, and functional hypotheses successively, which assists LLM in observing the evidence from trivial patterns to non-trivial patterns. Regarding hypothesis confirmation, we utilize Glymour's theory of bootstrap confirmation, a hypothesis confirmation theory from the philosophy of science that can confirm a set of hypotheses. We use its principles to confirm the object hypotheses, relational hypotheses, and functional hypotheses. Empirical studies on four inductive reasoning scenarios of different natures, involving causal induction, concept learning, grammar learning, and abstract reasoning, demonstrate that HypoBootstrap significantly outperforms existing methods.

Quantization and fine-tuning are crucial for deploying large language models (LLMs) on resource-constrained edge devices. However, fine-tuning quantized models presents significant challenges, primarily stemming from: First, the mismatch in data types between the low-precision quantized weights (e.g., 4-bit) and the high-precision adaptation weights (e.g., 16-bit). This mismatch limits the computational efficiency advantage offered by quantized weights during inference. Second, potential accuracy degradation when merging these high-precision adaptation weights into the low-precision quantized weights, as the adaptation weights often necessitate approximation or truncation. Third, as far as we know, no existing methods support the lossless merging of adaptation while adjusting all quantized weights. To address these challenges, we introduce lossless ternary adaptation for quantization-aware fine-tuning (LoTA-QAF). This is a novel fine-tuning method specifically designed for quantized LLMs, enabling the lossless merging of ternary adaptation weights into quantized weights and the adjustment of all quantized weights. LoTA-QAF operates through a combination of: i) A custom-designed ternary adaptation (TA) that aligns ternary weights with the quantization grid and uses these ternary weights to adjust quantized weights. ii) A TA-based mechanism that enables the lossless merging of adaptation weights. iii) Ternary signed gradient descent (t-SignSGD) for updating the TA weights. We apply LoTA-QAF to Llama-3.1/3.3 and Qwen-2.5 model families and validate its effectiveness on several downstream tasks. On the MMLU benchmark, our method effectively recovers performance for quantized models, surpassing 16-bit LoRA by up to 5.14\%. For task-specific fine-tuning, 16-bit LoRA achieves superior results, but LoTA-QAF still outperforms other methods. Code is available in github.com/KingdalfGoodman/LoTA-QAF.

Multimodal Large Language Models (MLLMs) have recently achieved remarkable progress in video understanding. However, their effectiveness in real-time streaming scenarios remains limited due to storage constraints of historical visual features and insufficient real-time spatiotemporal reasoning. To address these challenges, we propose StreamForest, a novel architecture specifically designed for streaming video understanding. Central to StreamForest is the Persistent Event Memory Forest, a memory mechanism that adaptively organizes video frames into multiple event-level tree structures. This process is guided by penalty functions based on temporal distance, content similarity, and merge frequency, enabling efficient long-term memory retention under limited computational resources. To enhance real-time perception, we introduce a Fine-grained Spatiotemporal Window, which captures detailed short-term visual cues to improve current scene perception. Additionally, we present OnlineIT, an instruction-tuning dataset tailored for streaming video tasks. OnlineIT significantly boosts MLLM performance in both real-time perception and future prediction. To evaluate generalization in practical applications, we introduce ODV-Bench, a new benchmark focused on real-time streaming video understanding in autonomous driving scenarios. Experimental results demonstrate that StreamForest achieves the state-of-the-art performance, with accuracies of 77.3% on StreamingBench, 60.5% on OVBench, and 55.6% on OVO-Bench. In particular, even under extreme visual token compression (limited to 1024 tokens), the model retains 96.8% of its average accuracy in eight benchmarks relative to the default setting. These results underscore the robustness, efficiency, and generalizability of StreamForest for streaming video understanding.

A single biomedical image can be segmented in multiple valid ways, depending on the application. For instance, a brain MRI may be divided according to tissue types, vascular territories, broad anatomical regions, fine-grained anatomy, or pathology. Existing automatic segmentation models typically either (1) support only a single protocol---the one they were trained on---or (2) require labor-intensive prompting to specify the desired segmentation. We introduce _Pancakes_, a framework that, given a new image from a previously unseen domain, automatically generates multi-label segmentation maps for _multiple_ plausible protocols, while maintaining semantic consistency across related images. In extensive experiments across seven previously unseen domains, _Pancakes_ consistently outperforms strong baselines, often by a wide margin, demonstrating its ability to produce diverse yet coherent segmentation maps on unseen domains.

Pruning is a widely used technique to compress large language models (LLMs) by removing unimportant weights, but it often suffers from significant performance degradation—especially under semi-structured sparsity constraints. Existing pruning methods primarily focus on estimating the importance of individual weights, which limits their ability to preserve critical capabilities of the model. In this work, we propose a new perspective: rather than merely selecting which weights to prune, we first redistribute parameter importance to make the model inherently more amenable to pruning. By minimizing the information entropy of normalized importance scores, our approach concentrates importance onto a smaller subset of weights, thereby enhancing pruning robustness. We instantiate this idea through DenoiseRotator, which applies learnable orthogonal transformations to the model’s weight matrices. Our method is model-agnostic and can be seamlessly integrated with existing pruning techniques such as Magnitude, SparseGPT, and Wanda. Evaluated on LLaMA3, Qwen2.5, and Mistral models under 50% unstructured and 2:4 semi-structured sparsity, DenoiseRotator consistently improves perplexity and zero-shot accuracy. For instance, on LLaMA3-70B pruned with SparseGPT at 2:4 semi-structured sparsity, DenoiseRotator reduces the perplexity gap to the dense model by 58%, narrowing the degradation from 8.1 to 3.4 points.

Egocentric video reasoning centers on an unobservable agent behind the camera who dynamically shapes the environment, requiring inference of hidden intentions and recognition of fine-grained interactions. This core challenge limits current multimodal large language models (MLLMs), which excel at visible event reasoning but lack embodied, first-person understanding. To bridge this gap, we introduce EgoThinker, a novel framework that endows MLLMs with robust egocentric reasoning capabilities through spatio-temporal chain-of-thought supervision and a two-stage learning curriculum. First, we introduce EgoRe-5M, a large-scale egocentric QA dataset constructed from 13M diverse egocentric video clips. This dataset features multi-minute segments annotated with detailed CoT rationales and dense hand–object grounding. Second, we employ SFT on EgoRe-5M to instill reasoning skills, followed by reinforcement fine-tuning (RFT) to further enhance spatio-temporal localization. Experimental results show that EgoThinker outperforms existing methods across multiple egocentric benchmarks, while achieving substantial improvements in fine-grained spatio-temporal localization tasks.

As large language models (LLMs) increasingly serve as close collaborators for humans, it is crucial that they express their reasoning in ways that humans can understand and learn from. However, this capability remains relatively less understood and under-evaluated. To address this, we introduce a conceptual framework for such Human-AI knowledge transfer capabilities and conduct the first large-scale user study (N=118) explicitly designed to measure it. In our two-phase setup, humans first ideate with an LLM on problem-solving strategies, then independently implement solutions, isolating the influence of model reasoning on human understanding. Our findings reveal that while model benchmark performance correlates with collaborative outcomes, this relationship is notably inconsistent with significant outliers, highlighting that knowledge transfer is a distinct capability requiring dedicated optimization. Our analysis uncovers behavioral and strategic factors that mediate successful knowledge transfer, and we release our code, dataset, and evaluation framework to support future work on communicatively aligned models.

Large language models (LLMs) have revolutionized natural language processing and are increasingly applied to other sequential data types, including genetic sequences. However, adapting LLMs to genetics presents significant challenges. Capturing complex genomic interactions requires modeling long-range global dependencies within DNA sequences, where interactions often span over 10,000 base pairs, even within a single gene. This poses substantial computational demands under conventional model architectures and training paradigms. Additionally, traditional LLM training approaches are suboptimal for DNA sequences: autoregressive training, while efficient for training, only supports unidirectional sequence understanding. However, DNA is inherently bidirectional. For instance, bidirectional promoters regulate gene expression in both directions and govern approximately 11% of human gene expression. Masked language models (MLMs) enable bidirectional understanding. However, they are inefficient since only masked tokens contribute to loss calculations at each training step. To address these limitations, we introduce JanusDNA, the first bidirectional DNA foundation model built upon a novel pretraining paradigm, integrating the optimization efficiency of autoregressive modeling with the bidirectional comprehension capability of masked modeling. JanusDNA's architecture leverages a Mamba-Attention Mixture-of-Experts (MoE) design, combining the global, high-resolution context awareness of attention mechanisms with the efficient sequential representation learning capabilities of Mamba. The MoE layers further enhance the model's capacity through sparse parameter scaling, while maintaining manageable computational costs. Notably, JanusDNA can process up to 1 million base pairs at single-nucleotide resolution on a single 80GB GPU using its hybrid architecture. Extensive experiments and ablation studies demonstrate that JanusDNA achieves new state-of-the-art performance on three genomic representation benchmarks. Remarkably, JanusDNA surpasses models with 250x more activated parameters, underscoring its efficiency and effectiveness. Code available at https://anonymous.4open.science/r/JanusDNA/.

Recent advances in large multimodal models have significantly advanced video comprehension, yet their performance remains limited in first-person scenarios. The interactive nature of egocentric videos is critical for applications like embodied intelligence, but introduces complex visual contexts that conventional models struggle to capture. To bridge this gap, we introduce OpenMMEgo with innovations across three dimensions: data, model, and training strategy. To provide rich spatiotemporal visual knowledge, we curate a large-scale, high-quality dataset named OME10M, comprising over 8.2M egocentric video QA pairs synthesized from Ego4D series. We also establish OMEBench, a comprehensive benchmark for rigorous egocentric understanding assessment. To alleviate the frequent viewpoint shifts inherent in egocentric videos, we implement semantic-aware visual token compression. Further, a curriculum learning strategy is complemented to foster stable learning across various data complexities. OpenMMEgo consistently improves the performance of LMMs on egocentric benchmarks without sacrificing general video understanding performance. Notably, Qwen2.5-VL tuned with OpenMMEgo substantially outperforms other models of the same size in egocentric video understanding. The data, weights and training code will be put at https://github.com/BeingBeyond/OpenMMEgo.

We study the sublinear multivariate mean estimation problem in $d$-dimensional Euclidean space. Specifically, we aim to find the mean $\mu$ of a ground point set $A$, which minimizes the sum of squared Euclidean distances of the points in $A$ to $\mu$. We first show that a multiplicative $(1+\varepsilon)$ approximation to $\mu$ can be found with probability $1-\delta$ using $O(\varepsilon^{-1}\log \delta^{-1})$ many independent uniform random samples, and provide a matching lower bound. Furthermore, we give two estimators with optimal sample complexity that can be computed in optimal running time for extracting a suitable approximate mean: 1. The coordinate-wise median of $\log \delta^{-1}$ sample means of sample size $\varepsilon^{-1}$. As a corollary, we also show improved convergence rates for this estimator for estimating means of multivariate distributions. 2. The geometric median of $\log \delta^{-1}$ sample means of sample size $\varepsilon^{-1}$. To compute a solution efficiently, we design a novel and simple gradient descent algorithm that is significantly faster for our specific setting than all other known algorithms for computing geometric medians. In addition, we propose an order statistics approach that is empirically competitive with these algorithms, has an optimal sample complexity and matches the running time up to lower order terms. We finally provide an extensive experimental evaluation among several estimators which concludes that the geometric-median-of-means-based approach is typically the most competitive in practice.

Process Reward Models (PRMs) have shown promise in enhancing the mathematical reasoning capabilities of Large Language Models (LLMs) through Test-Time Scaling (TTS). However, their integration into multimodal reasoning remains largely unexplored. In this work, we take the first step toward unlocking the potential of PRMs in multimodal mathematical reasoning. We identify three key challenges: (i) the scarcity of high-quality reasoning data constrains the capabilities of foundation Multimodal Large Language Models (MLLMs), which imposes further limitations on the upper bounds of TTS and reinforcement learning (RL); (ii) a lack of automated methods for process labeling within multimodal contexts persists; (iii) the employment of process rewards in unimodal RL faces issues like reward hacking, which may extend to multimodal scenarios. To address these issues, we introduce URSA, a three-stage Unfolding multimodal pRocess-Supervision Aided training framework. We first construct MMathCoT-1M, a high-quality large-scale multimodal Chain-of-Thought (CoT) reasoning dataset, to build a stronger math reasoning foundation MLLM, URSA-8B. Subsequently, we go through an automatic process to synthesize process supervision data, which emphasizes both logical correctness and perceptual consistency. We introduce DualMath-1.1M to facilitate the training of URSA-8B-RM. Finally, we propose Process-Supervised Group-Relative-Policy-Optimization (PS-GRPO), pioneering a multimodal PRM-aided online RL method that outperforms vanilla GRPO. With PS-GRPO application, URSA-8B-PS-GRPO outperforms Gemma3-12B and GPT-4o by 8.4% and 2.7% on average across 6 benchmarks.

Vision Transformers (ViTs) have become a universal backbone for both image recognition and image generation. Yet their Multi–Head Self–Attention (MHSA) layer still performs a quadratic query–key interaction for \emph{every} token pair, spending the bulk of computation on visually weak or redundant correlations. We introduce \emph{Visual–Contrast Attention} (VCA), a drop-in replacement for MHSA that injects an explicit notion of discrimination while reducing the theoretical complexity from $\mathcal{O}(N^{2}C)$ to $\mathcal{O}(N n C)$ with $n\!\ll\!N$. VCA first distils each head’s dense query field into a handful of spatially pooled \emph{visual–contrast tokens}, then splits them into a learnable \emph{positive} and \emph{negative} stream whose differential interaction highlights what truly separates one region from another. The module adds fewer than $0.3$\,M parameters to a DeiT-Tiny backbone, requires no extra FLOPs, and is wholly architecture-agnostic. Empirically, VCA lifts DeiT-Tiny top-1 accuracy on ImageNet-1K from $72.2\%$ to \textbf{$75.6\%$} (+$3.4$) and improves three strong hierarchical ViTs by up to $3.1$\%, while in class-conditional ImageNet generation it lowers FID-50K by $2.1$ to $5.2$ points across both diffusion (DiT) and flow (SiT) models. Extensive ablations confirm that (i) spatial pooling supplies low-variance global cues, (ii) dual positional embeddings are indispensable for contrastive reasoning, and (iii) combining the two in both stages yields the strongest synergy. VCA therefore offers a simple path towards faster and sharper Vision Transformers. The source code is available at \href{https://github.com/LeapLabTHU/LinearDiff}{https://github.com/LeapLabTHU/LinearDiff}.

Understanding cellular responses to chemical interventions is critical to the discovery of effective therapeutics. Because individual biological techniques often measure only one axis of cellular response at a time, high-quality multimodal datasets are needed to unlock a holistic understanding of how cells respond to treatments and to advance computational methods that integrate modalities. However, many techniques destroy cells and thus preclude paired measurements, and attempts to match disparate unimodal datasets are often confounded by data being generated in incompatible experimental settings. Here we introduce scGeneScope, a multimodal single‑cell RNA sequencing (scRNA-seq) and Cell Painting microscopy image dataset conditionally paired by chemical treatment, designed to facilitate the development and benchmarking of unimodal, multimodal, and multiple profile machine learning methods for cellular profiling. 28 chemicals, each acting on distinct biological pathways or mechanisms of action (MoAs), were applied to U2-OS cells in two experimental data generation rounds, creating paired sets of replicates that were then profiled independently by scRNA‑seq or Cell Painting. Using scGeneScope, we derive a replicate- and experiment-split treatment identification benchmark simulating MoA discovery under realistic laboratory variability conditions and evaluate unimodal, multimodal, and multiprofile models ranging in complexity from linear approaches to recent foundation models. Multiprofile integration improved performance in both the unimodal and multimodal settings, with gains more consistent in the former. Evaluation of unimodal models for MoA identification demonstrated that recent scRNA-seq foundation models deployed zero-shot were consistently outperformed by classic fit-to-data methods, underscoring the need for careful, realistic benchmarking in machine learning for biology. We release the scGeneScope dataset and benchmarking code to support further research.

The success of DeepSeek-R1 demonstrates the immense potential of using reinforcement learning (RL) to enhance LLMs' reasoning capabilities. This paper introduces Retrv-R1, the first R1-style MLLM specifically designed for multimodal universal retrieval, achieving higher performance by employing step-by-step reasoning to produce more accurate retrieval results. We find that directly applying the methods of DeepSeek-R1 to retrieval tasks is not feasible, mainly due to (1) the high computational cost caused by the large token consumption required for multiple candidates with reasoning processes, and (2) the instability and suboptimal results when directly applying RL to train for retrieval tasks. To address these issues, Retrv-R1 introduces an information compression module with a details inspection mechanism, which enhances computational efficiency by reducing the number of tokens while ensuring that critical information for challenging candidates is preserved. Additionally, a new training paradigm is proposed, including an activation stage using a retrieval-tailored synthetic CoT dataset for more effective optimization, followed by RL with a novel curriculum reward to improve both performance and efficiency. Incorporating these novel designs, Retrv-R1 achieves SOTA performance, high efficiency, and strong generalization ability, as demonstrated by extensive experiments across multiple benchmarks and tasks.

Transformer models have become the dominant backbone for sequence modeling, leveraging self-attention to produce contextualized token representations. These are typically aggregated into fixed-size vectors via pooling operations for downstream tasks. While much of the literature has focused on attention mechanisms, the role of pooling remains underexplored despite its critical impact on model behavior. In this paper, we introduce a theoretical framework that rigorously characterizes the expressivity of Transformer-based models equipped with widely used pooling methods by deriving closed-form bounds on their representational capacity and the ability to distinguish similar inputs. Our analysis extends to different variations of attention formulations, demonstrating that these bounds hold across diverse architectural variants. We empirically evaluate pooling strategies across tasks requiring both global and local contextual understanding, spanning three major modalities: computer vision, natural language processing, and time-series analysis. Results reveal consistent trends in how pooling choices affect accuracy, sensitivity, and optimization behavior. Our findings unify theoretical and empirical perspectives, providing practical guidance for selecting or designing pooling mechanisms suited to specific tasks. This work positions pooling as a key architectural component in Transformer models and lays the foundation for more principled model design beyond attention alone.

Large Multimodal Models (LMMs) have made significant breakthroughs with the advancement of instruction tuning. However, while existing models can understand images and videos at a holistic level, they still struggle with instance-level understanding that requires a more fine-grained comprehension and alignment. Instance-level understanding is crucial for LMMs, as it focuses on the specific elements that we are most interested in. Excitingly, existing works find that the state-of-the-art LMMs exhibit strong instance understanding capabilities when provided with explicit visual cues. Motivated by this, we proposed Inst-IT, a solution to enhance LMMs in Instance understanding via explicit visual prompt Instruction Tuning for instance guidance. Inst-IT consists of a benchmark to diagnose multimodal instance-level understanding, a large-scale instruction-tuning dataset, and a continuous instruction-tuning training paradigm to effectively enhance spatial-temporal instance understanding capabilities of existing LMMs. Experimental results show that, enhanced by Inst-IT, our models not only achieve outstanding performance on Inst-IT-Bench and other instance understanding benchmarks, but also demonstrate significant improvements across various generic image and video understanding benchmarks. This highlights that our method not only boosts instance-level understanding but also strengthens the overall capabilities of generic image and video comprehension.