NeurIPS 2025 Papers

Biological and artificial learners are inherently exposed to a stream of data and experience throughout their lifetimes and must constantly adapt to, learn from, or selectively ignore the ongoing input. Recent findings reveal that, even when the performance remains stable, the underlying neural representations can change gradually over time, a phenomenon known as representational drift. Studying the different sources of data and noise that may contribute to drift is essential for understanding lifelong learning in neural systems. However, a systematic study of drift across architectures and learning rules, and the connection to task, are missing. Here, in an online learning setup, we characterize drift as a function of data distribution, and specifically show that the learning noise induced by task-irrelevant stimuli, which the agent learns to ignore in a given context, can create long-term drift in the representation of task-relevant stimuli. Using theory and simulations, we demonstrate this phenomenon both in Hebbian-based learning---Oja's rule and Similarity Matching---and in stochastic gradient descent applied to autoencoders and a supervised two-layer network. We consistently observe that the drift rate increases with the variance and the dimension of the data in the task-irrelevant subspace. We further show that this yields different qualitative predictions for the geometry and dimension-dependency of drift than those arising from Gaussian synaptic noise. Overall, our study links the structure of stimuli, task, and learning rule to representational drift and could pave the way for using drift as a signal for uncovering underlying computation in the brain.

Transformers and their attention mechanism have been revolutionary in the field of Machine Learning. While originally proposed for the language data, they quickly found their way to the image, video, graph, etc. data modalities with various signal geometries. Despite this versatility, generalizing the attention mechanism to scenarios where data is presented at different scales from potentially different modalities is not straightforward. The attempts to incorporate hierarchy and multi-modality within transformers are largely based on ad hoc heuristics, which are not seamlessly generalizable to similar problems with potentially different structures. To address this problem, in this paper, we take a fundamentally different approach: we first propose a mathematical construct to represent multi-modal, multi-scale data. We then mathematically derive the neural attention mechanics for the proposed construct from the first principle of entropy minimization. We show that the derived formulation is optimal in the sense of being the closest to the standard Softmax attention while incorporating the inductive biases originating from the hierarchical/geometric information of the problem. We further propose an efficient algorithm based on dynamic programming to compute our derived attention mechanism. By incorporating it within transformers, we show that the proposed hierarchical attention mechanism not only can be employed to train transformer models in hierarchical/multi-modal settings from scratch, but it can also be used to inject hierarchical information into classical, pre-trained transformer models post training, resulting in more efficient models in zero-shot manner.

Multi-Agent Debate (MAD) has emerged as a promising paradigm for improving the performance of large language models through collaborative reasoning. Despite recent advances, the key factors driving MAD’s effectiveness remain unclear. In this work, we disentangle MAD into two key components–Majority Voting and inter-agent Debate–and assess their respective contributions. Through extensive experiments across seven NLP benchmarks, we find that Majority Voting alone accounts for most of the performance gains typically attributed to MAD. To explain this, we propose a theoretical framework that models debate as a stochastic process. We prove that it induces a martingale over agents’ belief trajectories, implying that debate alone does not improve expected correctness. Guided by these insights, we demonstrate that targeted interventions, by biasing the belief update toward correction, can meaningfully enhance debate effectiveness. Overall, our findings suggest that while MAD has potential, simple ensembling methods remain strong and more reliable alternatives in many practical settings. Code is released in https://github.com/deeplearning-wisc/debate-or-vote.

As the field progresses toward Artificial General Intelligence (AGI), there is a pressing need for more comprehensive and insightful evaluation frameworks that go beyond aggregate performance metrics. This paper introduces a unified rating system that jointly models the difficulty of individual test cases and the competency of AI models (or humans) across vision, language, and action domains. Unlike existing metrics that focus solely on models, our approach allows for fine-grained, difficulty-aware evaluations through competitive interactions between models and tasks, capturing both the long-tail distribution of real-world challenges and the competency gap between current models and full task mastery. We validate the generalizability and robustness of our system through extensive experiments on multiple established datasets and models across distinct AGI domains. The resulting rating distributions offer novel perspectives and interpretable insights into task difficulty, model progression, and the outstanding challenges that remain on the path to achieving full AGI task mastery. We have made our code and results publicly available at https://ss47816.github.io/AGI-Elo/.

This paper's primary objective is to develop a robust generalist perception model capable of addressing multiple tasks under constraints of computational resources and limited training data. We leverage text-to-image diffusion models pre-trained on billions of images and successfully introduce our DICEPTION, a visual generalist model. Exhaustive evaluations demonstrate that DICEPTION effectively tackles diverse perception tasks, even achieving performance comparable to SOTA single-task specialist models. Specifically, we achieve results on par with SAM-vit-h using only 0.06% of their data (e.g., 600K vs.\ 1B pixel-level annotated images). We designed comprehensive experiments on architectures and input paradigms, demonstrating that the key to successfully re-purposing a single diffusion model for multiple perception tasks lies in maximizing the preservation of the pre-trained model's prior knowledge. Consequently, DICEPTION can be trained with substantially lower computational costs than conventional models requiring training from scratch. Furthermore, adapting DICEPTION to novel tasks is highly efficient, necessitating fine-tuning on as few as 50 images and approximately 1% of its parameters. Finally, we demonstrate that a subtle application of classifier-free guidance can improve the model's performance on depth and normal estimation. We also show that pixel-aligned training, as is characteristic of perception tasks, significantly enhances the model's ability to preserve fine details. DICEPTION offers valuable insights and presents a promising direction for the development of advanced diffusion-based visual generalist models.

Comprehending natural language and following human instructions are critical capabilities for intelligent agents. However, the flexibility of linguistic instructions induces substantial ambiguity across language-conditioned tasks, severely degrading algorithmic performance. To address these limitations, we present a novel method named DAIL (Distributional Aligned Learning), featuring two key components: distributional policy and semantic alignment. Specifically, we provide theoretical results that the value distribution estimation mechanism enhances task differentiability. Meanwhile, the semantic alignment module captures the correspondence between trajectories and linguistic instructions. Extensive experimental results on both structured and visual observation benchmarks demonstrate that DAIL effectively resolves instruction ambiguities, achieving superior performance to baseline methods. Our implementation is available at https://github.com/RunpengXie/Distributional-Aligned-Learning.

Predicting future dynamics is crucial for applications like autonomous driving and robotics, where understanding the environment is key. Existing pixel-level methods are computationally expensive and often focus on irrelevant details. To address these challenges, we introduce DINO-Foresight, a novel framework that operates in the semantic feature space of pretrained Vision Foundation Models (VFMs). Our approach trains a masked feature transformer in a self-supervised manner to predict the evolution of VFM features over time. By forecasting these features, we can apply off-the-shelf, task-specific heads for various scene understanding tasks. In this framework, VFM features are treated as a latent space, to which different heads attach to perform specific tasks for future-frame analysis. Extensive experiments show the very strong performance, robustness and scalability of our framework.

Visual tokens consume substantial computational resources in multi-modal large models (MLLMs), significantly compromising their efficiency. Recent works have attempted to improve efficiency by compressing visual tokens during training, either through modifications to model components or by introducing additional parameters. However, they often overlook the increased learning difficulty caused by such compression, as the model’s parameter space struggles to quickly adapt to the substantial perturbations in the feature space induced by token compression. In this work, we propose to develop Efficient MLLMs via Progressive Consistency Distillation (EPIC), a progressive learning framework. Specifically, by decomposing the feature space perturbations introduced by token compression along the token-wise and layer-wise dimensions, we introduce token consistency distillation and layer consistency distillation, respectively, aiming to reduce the training difficulty by leveraging guidance from a teacher model and following a progressive learning trajectory. Extensive experiments demonstrate the superior effectiveness, robustness, and generalization capabilities of our proposed framework.

Adapting large-scale foundation models in multi-task scenarios often suffers from task conflict and oblivion. To mitigate such issues, we propose a novel "model MoE-ization" strategy that leads to a conflict- and oblivion-resistant multi-task adaptation method. Given a weight matrix of a pre-trained model, our method applies SVD to it and introduces a learnable router to adjust its singular values based on tasks and samples. Accordingly, the weight matrix becomes a Mixture of Orthogonal Rank-one Experts (MoORE), in which each expert corresponds to the outer product of a left singular vector and the corresponding right one. We can improve the model capacity by imposing a learnable orthogonal transform on the right singular vectors. Unlike low-rank adaptation (LoRA) and its MoE-driven variants, MoORE guarantees the experts' orthogonality and maintains the column space of the original weight matrix. These two properties make the adapted model resistant to the conflicts among the new tasks and the oblivion of its original tasks, respectively. Experiments on various datasets demonstrate that MoORE outperforms existing multi-task adaptation methods consistently, showing its superiority in terms of conflict- and oblivion-resistance. The code is available at https://github.com/DaShenZi721/MoORE.

Multi-source transfer learning provides an effective solution to data scarcity in real-world supervised learning scenarios by leveraging multiple source tasks. In this field, existing works typically use all available samples from sources in training, which constrains their training efficiency and may lead to suboptimal results. To address this, we propose a theoretical framework that answers the question: what is the optimal quantity of source samples needed from each source task to jointly train the target model? Specifically, we introduce a generalization error measure based on K-L divergence, and minimize it based on high-dimensional statistical analysis to determine the optimal transfer quantity for each source task. Additionally, we develop an architecture-agnostic and data-efficient algorithm OTQMS to implement our theoretical results for target model training in multi-source transfer learning. Experimental studies on diverse architectures and two real-world benchmark datasets show that our proposed algorithm significantly outperforms state-of-the-art approaches in both accuracy and data efficiency. The code is available at https://github.com/zqy0126/OTQMS.

Existing tracking-by-detection Multi-Object Tracking methods mainly rely on associating objects with tracklets using motion and appearance features. However, variations in viewpoint and occlusions can result in discrepancies between the features of current objects and those of historical tracklets. To tackle these challenges, this paper proposes a novel Spatial-Temporal Tracklet Graph Matching paradigm (STAR). The core idea of STAR is to achieve long-term, reliable object association through the association of ``tracklet clips (TCs)". TCs are segments of confidently associated multi-object trajectories, which are linked through graph matching. Specifically, STAR initializes TCs using a Confident Initial Tracklet Generator (CITG) and constructs a TC graph via Tracklet Clip Graph Construction (TCGC). In TCGC, each object in a TC is treated as a vertex, with the appearance and local topology features encoded on the vertex. The vertices and edges of the TC graph are then updated through message propagation to capture higher-order features. Finally, a Tracklet Clip Graph Matching (TCGM) method is proposed to efficiently and accurately associate the TCs through graph matching. STAR is model-agnostic, allowing for seamless integration with existing methods to enhance their performance. Extensive experiments on diverse datasets, including MOTChallenge, DanceTrack, and VisDrone2021-MOT, demonstrate the robustness and versatility of STAR, significantly improving tracking performance under challenging conditions.

While Transformer self-attention offers strong parallelism, the Key-Value (KV) cache grows linearly with sequence length and becomes a bottleneck for inference efficiency. Multi-head latent attention was recently developed to compress the KV cache into a low-rank latent space. This paper proposes Multi-head Temporal Latent Attention (MTLA), which further reduces the KV cache size along the temporal dimension, greatly lowering the memory footprint of self-attention inference. MTLA employs a hyper-network to dynamically merge temporally adjacent KV cache vectors. To address the mismatch between the compressed KV cache and processed sequence lengths, a stride-aware causal mask is proposed to ensure efficient parallel training and consistency with inference behaviour. Experiments across tasks, including speech translation, speech recognition, speech understanding and text summarisation, demonstrate that MTLA achieves competitive performance compared to standard Multi-Head Attention (MHA), while greatly improving inference speed and GPU memory usage. For example, on a English-German speech translation task, MTLA achieves a 5.3$\times$ speedup and a reduction in GPU memory usage by a factor of 8.3 compared to MHA, while maintaining translation quality.

While large language models (LLMs) can solve PhD-level reasoning problems over long context inputs, they still struggle with a seemingly simpler task: *following explicit length instructions*—e.g., *write a 10,000-word novel*. Additionally, models often generate far too short outputs, terminate prematurely, or even refuse the request. Existing benchmarks focus primarily on evaluating generations quality, but often overlook whether the generations meet length constraints. To this end, we introduce **Length Instruction Following Evaluation Benchmark** (LIFEBench) to comprehensively evaluate LLMs' ability to follow length instructions across diverse tasks and a wide range of specified lengths. LIFEBench consists of 10,800 instances across 4 task categories in both English and Chinese, covering length constraints ranging from 16 to 8192 words. We evaluate 26 widely-used LLMs and find that most models reasonably follow short-length instructions but deteriorate sharply beyond a certain threshold. Surprisingly, almost all models fail to reach the vendor-claimed maximum output lengths in practice, as further confirmed by our evaluations extending up to 32K words. Even long-context LLMs, despite their extended input-output windows, counterintuitively fail to improve length-instructions following. Notably, Reasoning LLMs outperform even specialized long-text generation models, achieving state-of-the-art length following. Overall, LIFEBench uncovers fundamental limitations in current LLMs' length instructions following ability, offering critical insights for future progress.

Task-based functional magnetic resonance imaging (fMRI) provides invaluable insights into human cognition but faces critical hurdles—low signal-to-noise ratio, high dimensionality, limited sample sizes, and costly data acquisition—that are amplified when integrating datasets across subjects or sites. This paper introduces orthogonal contrastive learning (OCL), a unified multi-representation framework for multi-subject fMRI analysis that aligns neural responses without requiring temporal preprocessing or uniform time-series lengths across subjects or sites. OCL employs two identical encoders: an online network trained with a contrastive loss that pulls together same-stimulus responses and pushes apart different-stimulus responses, and a target network whose weights track the online network via exponential moving average to stabilize learning. Each OCL network layer combines QR decomposition for orthogonal feature extraction, locality-sensitive hashing (LSH) to produce compact subject-specific signatures, positional encoding to embed temporal structure alongside spatial features, and a transformer encoder to generate discriminative, stimulus-aligned embeddings. We further enhance OCL with an unsupervised pretraining stage on fMRI-like synthetic data and demonstrate a transfer-learning workflow for multi-site studies. Across extensive experiments on multi-subject and multi-site fMRI benchmarks, OCL consistently outperforms state-of-the-art alignment and analysis methods in both representation quality and downstream classification accuracy.

Endoscopic procedures are essential for diagnosing and treating internal diseases, and multi-modal large language models (MLLMs) are increasingly applied to assist in endoscopy analysis. However, current benchmarks are limited, as they typically cover specific endoscopic scenarios and a small set of clinical tasks, failing to capture the real-world diversity of endoscopic scenarios and the full range of skills needed in clinical workflows. To address these issues, we introduce EndoBench, the first comprehensive benchmark specifically designed to assess MLLMs across the full spectrum of endoscopic practice with multi-dimensional capacities. EndoBench encompasses 4 distinct endoscopic scenarios, 12 specialized clinical tasks with 12 secondary subtasks, and 5 levels of visual prompting granularities, resulting in 6,832 rigorously validated VQA pairs from 21 diverse datasets. Our multi-dimensional evaluation framework mirrors the clinical workflow—spanning anatomical recognition, lesion analysis, spatial localization, and surgical operations—to holistically gauge the perceptual and diagnostic abilities of MLLMs in realistic scenarios. We benchmark 23 state-of-the-art models, including general-purpose, medical-specialized, and proprietary MLLMs, and establish human clinician performance as a reference standard. Our extensive experiments reveal: (1) proprietary MLLMs outperform open-source and medical-specialized models overall, but still trail human experts; (2) medical-domain supervised fine-tuning substantially boosts task-specific accuracy; and (3) model performance remains sensitive to prompt format and clinical task complexity. EndoBench establishes a new standard for evaluating and advancing MLLMs in endoscopy, highlighting both progress and persistent gaps between current models and expert clinical reasoning. We publicly release our benchmark and code.

Previous studies have compared neural activities in the visual cortex to representations in deep neural networks trained on image classification. Interestingly, while some suggest that their representations are highly similar, others argued the opposite. Here, we propose a new approach to characterize the similarity of the decision strategies of two observers (models or brains) using decision variable correlation (DVC). DVC quantifies the image-by-image correlation between the decoded decisions based on the internal neural representations in a classification task. Thus, it can capture task-relevant information rather than general representational alignment. We evaluate DVC using monkey V4/IT recordings and network models trained on image classification tasks. We find that model–model similarity is comparable to monkey-monkey similarity, whereas model–monkey similarity is consistently lower. Strikingly, DVC decreases with increasing network performance on ImageNet-1k. Adversarial training does not improve model–monkey similarity in task-relevant dimensions assessed using DVC, although it markedly increases the model–model similarity. Similarly, pre-training on larger datasets does not improve model–monkey similarity. These results suggest a divergence between the task-relevant representations in monkey V4/IT and those learned by models trained on image classification tasks.

Deep multi-modal clustering (DMC) aims to explore the correlated information from different modalities to improve the clustering performance. Most existing DMCs attempt to investigate the consistency or/and complementarity information by fusing all modalities, but this will lead to the following challenges: 1) Information conflicts between modalities emerge. 2) Information-rich modalities may be weakened. To address the above challenges, we propose a diversity-oriented deep multi-modal clustering (DDMC) method, where the core is dominant modality enhancement instead of multi-modal fusion. Specifically, we select the modality with the highest average silhouette coefficient as the dominant modality, then learn the diversity information between the dominant madality and the remaining ones with diversity learning, and finally enhance the dominant modality for clustering. Extensive experiments show the superiority of the proposed method over several compared DMC methods. To our knowledge, this is the first work to perform multi-modal clustering by enhancing the dominant modality instead of fusion.

Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks but require substantial computational resources, limiting their deployment in resource-constrained environments. While one-shot pruning methods can reduce model size without expensive retraining, they typically optimize for single objectives, ignoring LLMs' multi-faceted applications. We introduce Multi-Objective One-Shot Pruning (MOSP), which formulates LLM pruning as a multi-objective optimization problem. MOSP efficiently generates a Pareto set of pruned models representing different capability trade-offs, allowing users to select solutions aligned with their preferences. The proposed approach identifies share core support while enabling specialized support. Experiments across various LLMs and sparsity levels demonstrate MOSP's superior performance in navigating multi-objective trade-offs compared to baseline methods.

Rheumatoid arthritis (RA) is a common autoimmune disease that has been the focus of research in computer-aided diagnosis (CAD) and disease monitoring. In clinical settings, conventional radiography (CR) is widely used for the screening and evaluation of RA due to its low cost and accessibility. The wrist is a critical region for the diagnosis of RA. However, CAD research in this area remains limited, primarily due to the challenges in acquiring high-quality instance-level annotations. (i) The wrist comprises numerous small bones with narrow joint spaces, complex structures, and frequent overlaps, requiring detailed anatomical knowledge for accurate annotation. (ii) Disease progression in RA often leads to osteophyte, bone erosion (BE), and even bony ankylosis, which alter bone morphology and increase annotation difficulty, necessitating expertise in rheumatology. This work presents a multi-task dataset for wrist bone in CR, including two tasks: (i) wrist bone instance segmentation and (ii) Sharp/van der Heijde (SvdH) BE scoring, which is the first public resource for wrist bone instance segmentation. This dataset comprises 1048 wrist conventional radiographs of 388 patients from six medical centers, with pixel-level instance segmentation annotations for 618 images and SvdH BE scores for 800 images. This dataset can potentially support a wide range of research tasks related to RA, including joint space narrowing (JSN) progression quantification, BE detection, bone deformity evaluation, and osteophyte detection. It may also be applied to other wrist-related tasks, such as carpal bone fracture localization. We hope this dataset will significantly lower the barrier to research on wrist RA and accelerate progress in CAD research within the RA-related domain. Benchmark \& Code: https://github.com/YSongxiao/RAM-W600 Data \& Dataset Card: https://huggingface.co/datasets/TokyoTechMagicYang/RAM-W600