NeurIPS 2025 论文列表

One persistent challenge in LLM research is the development of attention mechanisms that are able to generalise from training on shorter contexts to inference on longer contexts. We propose two conditions that we expect all effective long-context attention mechanisms to have: scale-invariant total attention, and scale-invariant attention sparsity. Under a Gaussian assumption, we show that a simple position-dependent transformation of the attention logits is sufficient for these conditions to hold. Experimentally we find that the resulting scale-invariant attention scheme gives considerable benefits in terms of validation loss when zero-shot generalising from training on short contexts to validation on longer contexts, and is effective at long-context retrieval.

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.

Preference alignment in diffusion models has primarily focused on benign human preferences (e.g., aesthetic). In this paper, we propose a novel perspective: framing unrestricted adversarial example generation as a problem of aligning with adversary preferences. Unlike benign alignment, adversarial alignment involves two inherently conflicting preferences: visual consistency and attack effectiveness, which often lead to unstable optimization and reward hacking (e.g., reducing visual quality to improve attack success). To address this, we propose APA (Adversary Preferences Alignment), a two-stage framework that decouples conflicting preferences and optimizes each with differentiable rewards. In the first stage, APA fine-tunes LoRA to improve visual consistency using rule-based similarity reward. In the second stage, APA updates either the image latent or prompt embedding based on feedback from a substitute classifier, guided by trajectory-level and step-wise rewards. To enhance black-box transferability, we further incorporate a diffusion augmentation strategy. Experiments demonstrate that APA achieves significantly better attack transferability while maintaining high visual consistency, inspiring further research to approach adversarial attacks from an alignment perspective.

Graph Transformer (GT) has recently emerged as a promising neural network architecture for learning graph-structured data. However, its global attention mechanism with quadratic complexity concerning the graph scale prevents wider application to large graphs. Effectively representing graph information while ensuring learning efficiency remains challenging, as our analysis reveals that current GT designs targeting scalability still suffer from the computational bottleneck related to graph-scale operations. In this work, we tackle the GT scalability issue by proposing HubGT, a scalable Graph Transformer boosted by fully decoupled graph processing and simplified learning. HubGT represents the graph by a novel hierarchical scheme exploiting hub labels, which is shown to be more informative than plain adjacency by offering global connections while promoting locality, and is particularly suitable for handling complex graph patterns such as heterophily. We also design algorithms for efficiently constructing and querying the hub label hierarchy tailored for the GT attention training in scalable deployments. Notably, the precomputation and training processes of HubGT achieve complexities linear to the number of graph edges and nodes, respectively, while the training stage completely removes graph-related computations, leading to favorable mini-batch capability and GPU utilization. Extensive experiments demonstrate that HubGT is efficient in terms of computational enhancement and mini-batch capability over existing GT designs on large-scale benchmarks, while achieving top-tier effectiveness on both homophilous and heterophilous graphs.

Mechanistic interpretability research seeks to reverse-engineer large language models (LLMs) by uncovering the internal representations of concepts within their activations. Sparse Autoencoders (SAEs) have emerged as a valuable tool for disentangling polysemantic neurons into more monosemantic, interpretable features. However, recent work on automatic explanation generation for these features has faced challenges: explanations tend to be overly broad and fail to take polysemanticity into consideration. This work addresses these limitations by introducing a similarity-based strategy for sourcing close negative sentences that more effectively falsify generated explanations. Additionally, we propose a structured, component-based format for feature explanations and a tree-based, iterative explanation method that refines explanations. We demonstrate that our structured format and tree-based explainer improve explanation quality, while our similarity-based evaluation strategy exposes biases in existing interpretability methods. We also analyze the evolution of feature complexity and polysemanticity across LLM layers, offering new insights into information content within LLMs' residual streams.

Multimodal large language models (MLLMs) are essential for building general-purpose AI assistants; however, they pose increasing safety risks. How can we ensure safety alignment of MLLMs to prevent undesired behaviors? Going further, it is critical to explore how to fine-tune MLLMs to preserve capabilities while meeting safety constraints. Fundamentally, this challenge can be formulated as a min-max optimization problem. However, existing datasets have not yet disentangled single preference signals into explicit safety constraints, hindering systematic investigation in this direction. Moreover, it remains an open question whether such constraints can be effectively incorporated into the optimization process for multi-modal models. In this work, we present the first exploration of the Safe RLHF-V -- the first multimodal safety alignment framework. The framework consists of: (I) BeaverTails-V, the first open-source dataset featuring dual preference annotations for helpfulness and safety, supplemented with multi-level safety labels (minor, moderate, severe); (II) Beaver-Guard-V, a multi-level guardrail system to proactively defend against unsafe queries and adversarial attacks. Applying the guard model over five rounds of filtering and regeneration significantly enhances the precursor model’s overall safety by an average of 40.9%. (II) Based on dual preference, we initiate the first exploration of multi-modal safety alignment within a constrained optimization. Experimental results demonstrate that Safe RLHF effectively improves both model helpfulness and safety. Specifically, Safe RLHF-V enhances model safety by 34.2% and helpfulness by 34.3%.

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.

Finetuning provides a scalable and cost-effective means of customizing language models for specific tasks or response styles, with greater reliability than prompting or in-context learning. In contrast, the conventional wisdom is that injecting knowledge via finetuning results in brittle performance and poor generalization. We argue that the dichotomy of "task customization" (e.g., instruction tuning) and "knowledge injection" (e.g., teaching new facts) is a distinction without a difference. We instead identify concrete factors that explain the heterogeneous effectiveness observed with finetuning. To this end, we conduct a large-scale experimental study of finetuning the frontier Gemini v1.5 model family on a spectrum of datasets that are artificially engineered to interpolate between the strengths and failure modes of finetuning. Our findings indicate that question-answer training data formats provide much stronger knowledge generalization than document/article-style training data, numerical information can be harder for finetuning to retain than categorical information, and models struggle to apply finetuned knowledge during multi-step reasoning even when trained on similar examples---all factors that render ``knowledge injection'' to be especially difficult, even after controlling for considerations like data augmentation and information volume. On the other hand, our findings also indicate that it is not fundamentally more difficult to finetune information about a real-world event than information about writing style.

The issue-resolving task, where a model generates patches to fix real-world bugs, has emerged as a key benchmark for evaluating the capabilities of large language models (LLMs). While SWE-bench has become the dominant benchmark in this domain, it suffers from several limitations: it has not been updated since its release, is restricted to only 12 repositories, and relies heavily on manual effort for constructing test instances and setting up executable environments, significantly limiting its scalability. We present SWE-bench-Live, a live-updatable benchmark designed to address these limitations. SWE-bench-Live currently includes 1,890 tasks derived from real GitHub issues created since 2024, spanning 223 repositories. Each task is accompanied by a dedicated Docker image to ensure reproducible execution. Additionally, we introduce an automated curation pipeline that streamlines the entire process from instance creation to environment setup, removing manual bottlenecks and enabling scalability and continuous updates. We evaluate a range of state-of-the-art models and agent frameworks on SWE-bench-Live, offering detailed empirical insights into their real-world bug-fixing capabilities. By providing a fresh, diverse, and executable benchmark grounded in live repository activity, SWE-bench-Live supports reliable, large-scale assessment of code LLMs and code agents in realistic development settings.

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.

Recent works have studied fair resource allocation in social settings, where fairness is judged by the impact of allocation decisions rather than more traditional minimum or maximum thresholds on the allocations themselves. Our work significantly adds to this literature by developing continuous resource allocation strategies that adhere to *equality of impact*, a generalization of equality of opportunity. We derive methods to maximize total welfare across groups subject to minimal violation of equality of impact, in settings where the outcomes of allocations are unknown but have a diminishing marginal effect. While focused on a two-group setting, our study addresses a broader class of welfare dynamics than explored in prior work. Our contributions are threefold. First, we introduce *Equality of Impact (EoI)*, a fairness criterion defined via group-level impact functions. Second, we design an online algorithm for non-noisy settings that leverages the problem’s geometric structure and achieves constant cumulative fairness regret. Third, we extend this approach to noisy environments with a meta-algorithm and empirically demonstrate that our methods find fair allocations and perform competitively relative to representative baselines.

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.

Offline Learning from Observations (LfO) focuses on enabling agents to imitate expert behavior using datasets that contain only expert state trajectories and separate transition data with suboptimal actions. This setting is both practical and critical in real-world scenarios where direct environment interaction or access to expert action labels is costly, risky, or infeasible. Most existing LfO methods attempt to solve this problem through state or state-action occupancy matching. They typically rely on pretraining a discriminator to differentiate between expert and non-expert states, which could introduce errors and instability—especially when the discriminator is poorly trained. While recent discriminator-free methods have emerged, they generally require substantially more data, limiting their practicality in low-data regimes. In this paper, we propose IOSTOM ($\textit{Imitation from Observation via State Transition Occupancy Matching}$), a novel offline LfO algorithm designed to overcome these limitations. Our approach formulates a learning objective based on the joint state visitation distribution. A key distinction of IOSTOM is that it first excludes actions entirely from the training objective. Instead, we learn an $\textit{implicit policy}$ that models transition probabilities between states, resulting in a more compact and stable optimization problem. To recover the expert policy, we introduce an efficient action inference mechanism that $\textit{avoids training an inverse dynamics model}$. Extensive empirical evaluations across diverse offline LfO benchmarks show that IOSTOM substantially outperforms state-of-the-art methods, demonstrating both improved performance and data 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.

Neural networks are widely used for image–related tasks but typically demand considerable computing power. Once a network has been trained, however, its memory‑ and compute‑footprint can be reduced by compression. In this work, we focus on compression through tensorization and low‑rank representations. Whereas classical approaches search for a low‑rank approximation by minimizing an isotropic norm such as the Frobenius norm in weight‑space, we use data‑informed norms that measure the error in function space. Concretely, we minimize the change in the layer’s output distribution, which can be expressed as $\lVert (W - \widetilde{W}) \Sigma^{1/2}\rVert_F$ where $\Sigma^{1/2}$ is the square root of the covariance matrix of the layer’s input and $W$, $\widetilde{W}$ are the original and compressed weights. We propose new alternating least square algorithms for the two most common tensor decompositions (Tucker‑2 and CPD) that directly optimize the new norm. Unlike conventional compression pipelines, which almost always require post‑compression fine‑tuning, our data‑informed approach often achieves competitive accuracy without any fine‑tuning. We further show that the same covariance‑based norm can be transferred from one dataset to another with only a minor accuracy drop, enabling compression even when the original training dataset is unavailable. Experiments on several CNN architectures (ResNet‑18/50, and GoogLeNet) and datasets (ImageNet, FGVC‑Aircraft, Cifar10, and Cifar100) confirm the advantages of the proposed method.

Diffusion models have emerged as a powerful class of generative models across various modalities, including image, video, and audio synthesis. However, their deployment is often limited by significant inference latency, primarily due to the inherently sequential nature of the denoising process. While existing parallelization strategies attempt to accelerate inference by distributing computation across multiple devices, they typically incur high communication overhead, hindering deployment on commercial hardware. To address this challenge, we propose $\textbf{ParaStep}$, a novel parallelization method based on a reuse-then-predict mechanism that parallelizes diffusion inference by exploiting similarity between adjacent denoising steps. Unlike prior approaches that rely on layer-wise or stage-wise communication, ParaStep employs lightweight, step-wise communication, substantially reducing overhead. ParaStep achieves end-to-end speedups of up to $\textbf{3.88}$$\times$ on SVD, $\textbf{2.43}$$\times$ on CogVideoX-2b, and $\textbf{6.56}$$\times$ on AudioLDM2-large, while maintaining generation quality. These results highlight ParaStep as a scalable and communication-efficient solution for accelerating diffusion inference, particularly in bandwidth-constrained environments.

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.