NeurIPS 2025 论文列表

We propose Fast Long-range Adaptive Memory for Event (FLAME), a novel scalable architecture that combines neuro-inspired feature extraction with robust structured sequence modeling to efficiently process asynchronous and sparse event camera data. As a departure from conventional input encoding methods, FLAME presents Event Attention Layer, a novel feature extractor that leverages neuromorphic dynamics (Leaky Integrate-and-Fire (LIF)) to directly capture multi-timescale features from event streams. The feature extractor is integrates with a structured state-space model with a novel Event-Aware HiPPO (EA-HiPPO) mechanism that dynamically adapts memory retention based on inter-event intervals to understand relationship across varying temporal scales and event sequences. A Normal Plus Low Rank (NPLR) decomposition reduces the computational complexity of state update from $\mathcal{O}(N^2)$ to $\mathcal{O}(Nr)$, where $N$ represents the dimension of the core state vector and $r$ is the rank of a low-rank component (with $r \ll N$). FLAME demonstrates state-of-the-art accuracy for event-by-event processing on complex event camera datasets.

Real-world applications are stretching context windows to hundreds of thousand of tokens while Large Language Models (LLMs) swell from billions to trillions of parameters. This dual expansion send compute and memory costs skyrocketing, making $\textit{token compression}$ indispensable. We introduce Vision Centric Token Compression ($\textbf{Vist}$), a $\textit{slow–fast}$ compression framework that mirrors human reading: the $\textit{fast}$ path renders distant tokens into images, letting a $\textbf{frozen, lightweight vision encoder}$ skim the low-salience context; the $\textit{slow}$ path feeds the proximal window into the LLM for fine-grained reasoning. A Probability-Informed Visual Enhancement (PVE) objective masks high-frequency tokens during training, steering the Resampler to concentrate on semantically rich regions—just as skilled reader gloss over function words. On eleven in-context learning benchmarks, $\textbf{Vist}$ achieves the same accuracy with 2.3$\times$ fewer tokens, cutting FLOPs by 16\% and memory by 50\%. This method delivers remarkable results, outperforming the strongest text encoder-based compression method CEPE by $\textbf{7.6}$\% on average over benchmarks like TriviaQA, NQ, PopQA, NLUI, and CLIN, setting a new standard for token efficiency in LLMs. The source code will be released.

Establishing object-level correspondence between egocentric and exocentric views is essential for intelligent assistants to deliver precise and intuitive visual guidance. However, this task faces numerous challenges, including extreme viewpoint variations, occlusions, and the presence of small objects. Existing approaches usually borrow solutions from video object segmentation models, but still suffer from the aforementioned challenges. Recently, the Segment Anything Model 2 (SAM 2) has shown strong generalization capabilities and excellent performance in video object segmentation. Yet, when simply applied to the ego-exo correspondence (EEC) task, SAM 2 encounters severe difficulties due to ineffective ego-exo feature fusion and limited long-term memory capacity, especially for long videos. Addressing these problems, we propose a novel EEC framework based on SAM 2 with long-term memories by presenting a dual-memory architecture and an adaptive feature routing module inspired by Mixture-of-Experts (MoE). Compared to SAM 2, our approach features **(i)** a Memory-View MoE module which consists of a dual-branch routing mechanism to adaptively assign contribution weights to each expert feature along both channel and spatial dimensions, and **(ii)** a dual-memory bank system with a simple yet effective compression strategy to retain critical long-term information while eliminating redundancy. In the extensive experiments on the challenging EgoExo4D benchmark, our method, dubbed ***LM-EEC***, achieves new state-of-the-art results and significantly outperforms existing methods and the SAM 2 baseline, showcasing its strong generalization across diverse scenarios. Our code and model are available at https://github.com/juneyeeHu/LM-EEC.

A transformer is merely a stack of learned data–to–data maps—yet those maps can hide rich algorithms. We train a linear, attention-only transformer on millions of masked-block completion tasks: each prompt is a masked low-rank matrix whose missing block may be (i) a scalar prediction target or (ii) an unseen kernel slice for Nyström extrapolation. The model sees only input–output pairs and a mean-squared loss; it is given no normal equations, no handcrafted iterations, and no hint that the tasks are related. Surprisingly, after training, algebraic unrolling reveals the same parameter-free update rule across all three resource regimes (full visibility, bandwidth-limited heads, rank-limited attention). We prove that this rule achieves second-order convergence on full-batch problems, cuts distributed iteration complexity, and remains accurate with compute-limited attention. Thus, a transformer trained solely to patch missing blocks implicitly discovers a unified, resource-adaptive iterative solver spanning prediction, estimation, and Nyström extrapolation—highlighting a powerful capability of in-context learning.

Efficient long-context modeling remains a critical challenge for natural language processing (NLP), as the time complexity of the predominant Transformer architecture scales quadratically with the sequence length. While state-space models (SSMs) offer alternative sub-quadratic solutions, they struggle to capture long-range dependencies effectively. In this work, we focus on analyzing and improving the long-context modeling capabilities of SSMs. We show that the widely used synthetic task, associative recall, which requires a model to recall a value associated with a single key without context, insufficiently represents the complexities of real-world long-context modeling. To address this limitation, we extend the associative recall to a novel synthetic task, joint recall, which requires a model to recall the value associated with a key given in a specified context. Theoretically, we prove that SSMs do not have the expressiveness to solve multi-query joint recall in sub-quadratic time complexity. To resolve this issue, we propose a solution based on integrating SSMs with Context-Dependent Sparse Attention (CDSA), which has the expressiveness to solve multi-query joint recall with sub-quadratic computation. To bridge the gap between theoretical analysis and real-world applications, we propose locality-sensitive Hashing Attention with sparse Key Selection (HAX), which instantiates the theoretical solution and is further tailored to natural language domains. Extensive experiments on both synthetic and real-world long-context benchmarks show that HAX consistently outperforms SSM baselines and SSMs integrated with context-independent sparse attention (CISA). Our code is available at: https://github.com/DeepGraphLearning/HAX.

Recent advancements in LLM pretraining have featured ever-expanding context windows to process longer sequences. However, our controlled study reveals that models pretrained with shorter context windows consistently outperform their long-context counterparts under a fixed token budget. This finding motivates us to explore an optimal context window scheduling strategy to better balance long-context capability with pretraining efficiency. To this end, we propose SkyLadder, a simple yet effective approach that implements a short-to-long context window transition. SkyLadder preserves strong standard benchmark performance, while matching or exceeding baseline results on long-context tasks. Through extensive experiments, we pretrain 1B-parameter models (up to 32K context) and 3B-parameter models (8K context) on 100B tokens, demonstrating that SkyLadder yields consistent gains of up to 3.7% on common benchmarks, while achieving up to 22% faster training speeds compared to baselines.

Large language models (LLMs) exhibit impressive in‑context learning (ICL) capabilities, yet the quality of their predictions is fundamentally limited by the few costly labeled demonstrations that can fit into a prompt. Meanwhile, there exist vast and continuously growing amounts of unlabeled data that may be closely related to the ICL task. How to utilize such unlabeled data to provably enhance the performance of ICL thus becomes an emerging fundamental question. In this work, we propose a novel augmented ICL framework, in which the prompt includes a small set of labeled examples alongside a block of unlabeled inputs. We focus on the multi-class linear classification setting and demonstrate that, with chain-of-thought (CoT) prompting, a multi-layer transformer can effectively emulate an expectation–maximization (EM) algorithm. This enables the transformer to implicitly extract useful information from both labeled and unlabeled data, leading to provable improvements in ICL accuracy. Moreover, we show that such a transformer can be trained via teacher forcing, with its parameters converging to the desired solution at a linear rate. Experiments demonstrate that the augmented ICL framework consistently outperforms conventional few-shot ICL, providing empirical support for our theoretical findings. To the best of our knowledge, this is the first theoretical study on the impact of unlabeled data on the ICL performance of transformers.

We introduce Eagle2.5, a frontier vision-language model (VLM) for long-context multimodal learning. Our work addresses the challenges in long video comprehension and high-resolution image understanding, introducing a generalist framework for both tasks. The proposed training framework incorporates Automatic Degrade Sampling and Image Area Preservation, two techniques that preserve contextual integrity and visual details. The framework also includes numerous efficiency optimizations in the pipeline for long-context data training. Finally, we propose Eagle-Video-110K, a novel dataset that integrates both story-level and clip-level annotations, facilitating long-video understanding. Eagle2.5 demonstrates substantial improvements on long-context multimodal benchmarks, providing a robust solution to the limitations of existing VLMs. Notably, our best model Eagle2.5-8B achieves 72.4\% on Video-MME with 512 input frames, matching the results of top-tier commercial model such as GPT-4o and large-scale open-source models like Qwen2.5-VL-72B and InternVL2.5-78B.

Long reasoning models (LRMs) such as OpenAI's o1 and DeepSeek's R1 have demonstrated remarkable abilities in complex reasoning tasks by scaling test-time compute and exhibiting human-like deep thinking. However, we identify a phenomenon we term underthinking, where LRMs frequently switch between different reasoning thoughts without sufficiently exploring promising paths to reach a correct solution. This behavior leads to inadequate depth of reasoning and decreased performance, particularly on challenging mathematical problems. To systematically analyze this issue, we conduct experiments on three challenging test sets and two representative open-source LRMs, revealing that frequent thought switching correlates with incorrect responses. We introduce a novel metric to quantify underthinking by measuring token efficiency in incorrect answers. To address underthinking, we propose a decoding strategy with thought switching penalty (Tip) that discourages premature transitions between thoughts, encouraging deeper exploration of each reasoning path. Experimental results demonstrate that our approach improves accuracy across challenging datasets without requiring model fine-tuning. Our findings contribute to understanding reasoning inefficiencies in LRMs and offer a practical solution to enhance their problem-solving capabilities. Our code is open-source and available at https://github.com/wangyuenlp/underthinking.

Long-context supervised fine-tuning (Long-SFT) plays a vital role in enhancing the performance of large language models (LLMs) on long-context tasks. To smoothly adapt LLMs to long-context scenarios, this process typically entails training on mixed datasets containing both long and short sequences. However, this heterogeneous sequence length distribution poses significant challenges for existing training systems, as they fail to simultaneously achieve high training efficiency for both long and short sequences, resulting in sub-optimal end-to-end system performance in Long-SFT. In this paper, we present a novel perspective on data scheduling to address the challenges posed by the heterogeneous data distributions in Long-SFT. We propose Skrull, a dynamic data scheduler specifically designed for efficient long-SFT. Through dynamic data scheduling, Skrull balances the computation requirements of long and short sequences, improving overall training efficiency. Furthermore, we formulate the scheduling process as a joint optimization problem and thoroughly analyze the trade-offs involved. Based on those analysis, Skrull employs a lightweight scheduling algorithm to achieve near-zero cost online scheduling in Long-SFT. Finally, we implement Skrull upon DeepSpeed, a state-of-the-art distributed training system for LLMs. Experimental results demonstrate that Skrull outperforms DeepSpeed by 3.76x on average (up to 7.54x) in real-world long-SFT scenarios.

Numerous studies have demonstrated that the Transformer architecture possesses the capability for in-context learning (ICL). In scenarios involving function approximation, context can serve as a control parameter for the model, endowing it with the universal approximation property (UAP). In practice, context is represented by tokens from a finite set, referred to as a vocabulary, which is the case considered in this paper, i.e., vocabulary in-context learning (VICL). We demonstrate that VICL in single-layer Transformers, without positional encoding, does not possess the UAP; however, it is possible to achieve the UAP when positional encoding is included. Several sufficient conditions for the positional encoding are provided. Our findings reveal the benefits of positional encoding from an approximation theory perspective in the context of in-context learning.

Diffusion models have achieved impressive performance in generating high-quality and diverse synthetic data. However, their success typically assumes a class-balanced training distribution. In real-world settings, multi-class data often follow a long-tailed distribution, where standard diffusion models struggle—producing low-diversity and lower-quality samples for underrepresented (tail) classes. While this degradation is well-documented, its underlying cause remains poorly understood. In this work, we investigate the behavior of diffusion models trained on long-tailed datasets and identify a key issue: the latent representations (from the bottleneck layer of the U-Net) for tail class subspaces exhibit significant overlap with those of head classes, leading to feature borrowing and poor generation quality. Importantly, we show that this is not merely due to limited data per class, but that the relative class imbalance significantly contributes to this phenomenon. To address this, we propose **CO**ntrastive **R**egularization for **A**ligning **L**atents (CORAL), a contrastive latent alignment framework that leverages supervised contrastive losses to encourage well-separated latent class representations. Experiments demonstrate that CORAL significantly improves both the diversity and visual quality of samples generated for tail classes relative to state-of-the-art methods.

As large language models increasingly gain popularity in real-world applications, processing extremely long contexts, often exceeding the model’s pre-trained context limits, has emerged as a critical challenge. While existing approaches to efficient long-context processing show promise, recurrent compression-based methods struggle with information preservation, whereas random access approaches require substantial memory resources. We introduce REFORM, a novel inference framework that efficiently handles long contexts through a two-phase approach. First, it incrementally processes input chunks while maintaining a compressed KV cache, constructs cross-layer context embeddings, and utilizes early exit strategy for improved efficiency. Second, it identifies and gathers essential tokens via similarity matching and selectively recomputes the KV cache. Compared to baselines, REFORM achieves over 50% and 27% performance gains on RULER and BABILong respectively at 1M context length. It also outperforms baselines on ∞-Bench, RepoEval, and MM-NIAH, demonstrating flexibility across diverse tasks and domains. Additionally, REFORM reduces inference time by 30% and peak memory usage by 5%, achieving both efficiency and superior performance.

The imbalance (or long-tail) is the nature of many real-world data distributions, which often induces the undesirable bias of deep classification models toward frequent classes, resulting in poor performance for tail classes. In this paper, we propose a novel two-stage learning approach to mitigate such a majority-biased tendency while preserving valuable information within datasets. Specifically, the first stage proposes a new representation learning technique from the information theory perspective. This approach is theoretically equivalent to minimizing intra-class distance, yielding an effective and well-separated feature space. The second stage develops a novel sampling strategy that selects mathematically informative instances, able to rectify majority-biased decision boundaries without compromising a model’s overall performance. As a result, our approach achieves the state-of-the-art performance across various long-tailed benchmark datasets, validated via extensive experiments. Our code is available at https://github.com/fudong03/BNS_IPDPP.

Long-context large language models (LLMs) have achieved remarkable advancements, driven by techniques like Rotary Position Embedding (RoPE) (Su et al., 2023) and its extensions (Chen et al., 2023; Liu et al., 2024c; Peng et al., 2023). By adjusting RoPE parameters and incorporating training data with extended contexts, we can train performant models with considerably longer input sequences. However, existing RoPE-based methods exhibit performance limitations when applied to extended context lengths. This paper presents a comprehensive analysis of various attention mechanisms, including RoPE, No Positional Embedding (NoPE), and Query-Key Normalization (QK-Norm), identifying their strengths and shortcomings in long-context modeling. Our investigation identifies distinctive attention patterns in these methods and highlights their impact on long-context performance, providing valuable insights for architectural design. on long context performance, providing valuable insights for architectural design. Building on these findings, we propose a novel architecture featuring a hybrid attention mechanism that integrates global and local attention spans. This design not only surpasses conventional RoPE-based transformer models with full attention in both long and short context tasks but also delivers substantial efficiency gains during training and inference.

The feedback that AI systems (e.g., recommender systems, chatbots) collect from user interactions is a crucial source of training data. While short-term feedback (e.g., clicks, engagement) is widely used for training, there is ample evidence that optimizing short-term feedback does not necessarily achieve the desired long-term objectives. Unfortunately, directly optimizing for long-term objectives is challenging, and we identify the disconnect in the timescales of short-term interventions (e.g., rankings) and the long-term feedback (e.g., user retention) as one of the key obstacles. To overcome this disconnect, we introduce the framework of MultiScale Policy Learning to contextually reconcile that AI systems need to act and optimize feedback at multiple interdependent timescales. Following a PAC-Bayes motivation, we show how the lower timescales with more plentiful data can provide a data-dependent hierarchical prior for faster learning at higher scales, where data is more scarce. As a result, the policies at all levels effectively optimize for the long-term. We instantiate the framework with MultiScale Off-Policy Bandit Learning (MSBL) and demonstrate its effectiveness on three tasks relating to recommender and conversational systems.

Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to handle actual production workloads with long-context requirements, demonstrating significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.

Complex, temporally evolving phenomena, from climate to brain activity, are governed by dynamical systems (DS). DS reconstruction (DSR) seeks to infer generative surrogate models of these from observed data, reproducing their long-term behavior. Existing DSR approaches require purpose-training for any new system observed, lacking the zero-shot and in-context inference capabilities known from LLMs. Here we introduce *DynaMix*, a novel multivariate ALRNN-based mixture-of-experts architecture pre-trained for DSR, the first DSR model able to generalize zero-shot to out-of-domain DS. Just from a provided context signal, without any re-training, DynaMix faithfully forecasts the long-term evolution of novel DS where existing time series (TS) foundation models, like Chronos, fail -- at a fraction of the number of parameters (0.1%) and orders of magnitude faster inference times. DynaMix outperforms TS foundation models in terms of long-term statistics, and often also short-term forecasts, even on real-world time series, like traffic or weather data, typically used for training and evaluating TS models, *but not at all part of DynaMix' training corpus*. We illustrate some of the failure modes of TS models for DSR problems, and conclude that models built on DS principles may bear a huge potential also for advancing the TS prediction field.

We prove that any Turing machine running on inputs of arbitrary length can be simulated by a constant bit-size transformer, as long as the context window is sufficiently long. This improves previous works, which require scaling up either the model's precision or the number of parameters on longer inputs. Furthermore, we prove that the complexity class SPACE$[s(n)]$ exactly characterizes the expressive power of a constant bit-size transformer with a context window of length $s(n)$. Our approach relies on simulating Post machines, a Turing-complete computational model. Post machines can be modeled as automata equipped with a queue, exhibiting computational behaviors naturally aligned with those of transformers. The behavioral similarity between transformers and Post machines may offer new insights into the mechanisms underlying the reasoning abilities of transformers.