NeurIPS 2025 论文列表

Training Transformer models on long sequences in a distributed setting poses significant challenges in terms of efficiency and scalability. Current methods are either constrained by the number of attention heads or excessive communication overheads. To address this problem, we propose StarTrail, a multi-dimensional concentric distributed training system for long sequences, fostering an efficient communication paradigm and providing additional tuning flexibility for communication arrangements. Specifically, StarTrail introduces an extra parallel dimension and divides the peer-to-peer communication into sub-rings to substantially reduce communication volume and avoid bandwidth bottlenecks. Through comprehensive experiments across diverse hardware environments and on both Natural Language Processing (NLP) and Computer Vision (CV) tasks, we demonstrate that our approach significantly surpasses state-of-the-art methods that support Long sequence lengths, achieving performance improvements of up to 77.12% on GPT-style models and up to 114.33% on DiT (Diffusion Transformer) models without affecting the computations results.

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.

Attention is the cornerstone of modern Large Language Models (LLMs). Yet its quadratic complexity hinders efficiency and scalability, especially for long-context processing. A promising approach is to leverage sparsity in attention. However, existing sparsity-based solutions predominantly rely on predefined patterns or heuristics at the attention head level, struggling to adapt dynamically to different contexts efficiently. We propose SeerAttention, a simple yet effective attention mechanism that directly learns the block-level attention sparsity from the LLM itself. Inspired by the gating mechanism in Mixture of Experts (MoE), SeerAttention augments the conventional attention with a **learnable gate** that **selectively activates important blocks** within the attention map. Specifically, the gate first pools the query (Q) and key (K) tensors along the sequence dimension and processes them through learnable linear layers. The resulting matrices are then multiplied together to produce the gating scores, which are used to predict block-level attention sparsity. Combined with our block-sparse FlashAttention kernel, SeerAttention can achieve significant speedup on GPUs. When applied to pre-trained LLMs, SeerAttention only requires training the gate parameters in a lightweight self-distillation manner, allowing rapid convergence. Our evaluation results demonstrate that SeerAttention achieves better model accuracy and lower latency for long-context pre-filling compared to prior methods. Code is available at: https://github.com/microsoft/SeerAttention.

As the length of input text grows, the key-value (KV) cache in LLMs imposes prohibitive GPU memory costs and limits long‐context inference on resource‐constrained devices. Existing approaches, such as KV quantization and pruning, reduce memory usage but suffer from numerical precision loss or suboptimal retention of key-value pairs. We introduce Low Rank Query and Key attention (LRQK), a two‐stage framework that jointly decomposes the full‐precision query and key matrices into compact rank-\(r\) factors during the prefill stage, and then uses these low-dimensional projections to compute proxy attention scores in \(\mathcal{O}(lr)\) time at each decode step. By selecting only the top-\(k\) tokens and a small fixed set of recent tokens, LRQK employs a mixed GPU–CPU cache with a hit-and-miss mechanism that transfers only missing full-precision KV pairs, thereby preserving exact attention outputs while reducing CPU–GPU data movement. Extensive experiments on the RULER and LongBench benchmarks with LLaMA-3-8B and Qwen2.5-7B demonstrate that LRQK matches or surpasses leading sparse-attention methods in long context settings, while delivering significant memory savings with minimal loss in accuracy. Our code is available at \url{https://github.com/tenghuilee/LRQK}.

State-space models (SSMs), particularly Mamba, emerge as an efficient Transformer alternative with linear complexity for long-sequence modeling. Recent empirical works demonstrate Mamba's in-context learning (ICL) capabilities competitive with Transformers, a critical capacity for large foundation models. However, theoretical understanding of Mamba’s ICL remains limited, restricting deeper insights into its underlying mechanisms. Even fundamental tasks such as linear regression ICL, widely studied as a standard theoretical benchmark for Transformers, have not been thoroughly analyzed in the context of Mamba. To address this gap, we study the training dynamics of Mamba on the linear regression ICL task. By developing novel techniques tackling non-convex optimization with gradient descent related to Mamba's structure, we establish an exponential convergence rate to ICL solution, and derive a loss bound that is comparable to Transformer's. Importantly, our results reveal that Mamba can perform a variant of \textit{online gradient descent} to learn the latent function in context. This mechanism is different from that of Transformer, which is typically understood to achieve ICL through gradient descent emulation. The theoretical results are verified by experimental simulation.

Transformer-based large language models (LLMs) cache context as key-value (KV) pairs during inference. As context length grows, KV cache sizes expand, leading to substantial memory overhead and increased attention latency. This paper introduces \textit{KVzip}, a query-agnostic KV cache eviction method enabling effective reuse of compressed KV caches across diverse queries. KVzip quantifies the importance of a KV pair using the underlying LLM to reconstruct original contexts from cached KV pairs, subsequently evicting pairs with lower importance. Extensive empirical evaluations demonstrate that KVzip reduces KV cache size by $3$-$4\times$ and FlashAttention decoding latency by approximately $2\times$, with negligible performance loss in question-answering, retrieval, reasoning, and code comprehension tasks. Evaluations include various models such as LLaMA3.1, Qwen2.5, and Gemma3, with context lengths reaching up to 170K tokens. KVzip significantly outperforms existing query-aware KV eviction methods, which suffer from performance degradation even at a 90\% cache budget ratio under multi-query scenarios.

Deep generative models, particularly diffusion models, have achieved remarkable success but face significant challenges when trained on real-world, long-tailed datasets- where few "head" classes dominate and many "tail" classes are underrepresented. This paper develops a theoretical framework for long-tailed learning via diffusion models through the lens of deep mutual learning. We introduce a novel regularized training objective that combines the standard diffusion loss with a mutual learning term, enabling balanced performance across all class labels, including the underrepresented tails. Our approach to learn via the proposed regularized objective is to formulate it as a multi-player game, with Nash equilibrium serving as the solution concept. We derive a non-asymptotic first-order convergence result for individual gradient descent algorithm to find the Nash equilibrium. We show that the Nash gap of the score network obtained from the algorithm is upper bounded by $\mathcal{O}(\frac{1}{\sqrt{T_{train}}}+\beta)$ where $\beta$ is the regularizing parameter and $T_{train}$ is the number of iterations of the training algorithm. Furthermore, we theoretically establish hyper-parameters for training and sampling algorithm that ensure that we find conditional score networks (under our model) with a worst case sampling error $\mathcal{O}(\epsilon+1), \forall \epsilon>0$ across all class labels. Our results offer insights and guarantees for training diffusion models on imbalanced, long-tailed data, with implications for fairness, privacy, and generalization in real-world generative modeling scenarios.

Inference-time computation has emerged as a promising scaling axis for improving large language model reasoning. However, despite yielding impressive performance, the optimal allocation of inference-time computation remains poorly understood. A central question is whether to prioritize sequential scaling (e.g., longer chains of thought) or parallel scaling (e.g., majority voting across multiple short chains of thought). In this work, we seek to illuminate the landscape of test-time scaling by demonstrating the existence of reasoning settings where sequential scaling offers an exponential advantage over parallel scaling. These settings are based on graph connectivity problems in challenging distributions of graphs. We validate our theoretical findings with comprehensive experiments across a range of language models, including models trained from scratch for graph connectivity with different chain of thought strategies as well as large reasoning models.

With the development of large language models (LLMs), efficient inference through Key-Value (KV) cache compression has attracted considerable attention, especially for long-context generation. To compress the KV cache, recent methods identify critical KV tokens through static modeling of attention scores. However, these methods often struggle to accurately determine critical tokens as they neglect the *temporal patterns* in attention scores, resulting in a noticeable degradation in LLM performance. To address this challenge, we propose **AttentionPredictor**, which is the **first learning-based method to directly predict attention patterns for KV cache compression and critical token identification**. Specifically, AttentionPredictor learns a lightweight, unified convolution model to dynamically capture spatiotemporal patterns and predict the next-token attention scores. An appealing feature of AttentionPredictor is that it accurately predicts the attention score and shares the unified prediction model, which consumes negligible memory, among all transformer layers. Moreover, we propose a cross-token critical cache prefetching framework that hides the token estimation time overhead to accelerate the decoding stage. By retaining most of the attention information, AttentionPredictor achieves **13$\times$** KV cache compression and **5.6$\times$** speedup in a cache offloading scenario with comparable LLM performance, significantly outperforming the state-of-the-arts. The code is available at https://github.com/MIRALab-USTC/LLM-AttentionPredictor.

Long-term time series forecasting (LTSF) aims to predict future trends based on historical data. While longer lookback windows theoretically offer more comprehensive insights, Transformer-based models often struggle with them. On one hand, longer windows introduce more noise and redundancy, hindering the model's learning process. On the other hand, Transformers suffer from attention dispersion and are prone to overfitting to noise, especially when processing long sequences. In this paper, we introduce the Maximum Effective Window (MEW) metric to assess a model's ability to effectively utilize the lookback window. We also propose two model-agnostic modules to enhance MEW, enabling models to better leverage historical data for improved performance. Specifically, to reduce redundancy and noise, we introduce the Information Bottleneck Filter (IBF), which employs information bottleneck theory to extract the most essential subsequences from the input. Additionally, we propose the Hybrid-Transformer-Mamba (HTM), which incorporates the Mamba mechanism for selective forgetting of long sequences while harnessing the Transformer's strong modeling capabilities for shorter sequences. We integrate these two modules into various Transformer-based models, and experimental results show that they effectively enhance MEW, leading to improved overall performance. Our code is available at \url{https://github.com/forever-ly/PIH}.

Large Language Models (LLMs) achieve impressive reasoning capabilities at the cost of substantial inference overhead, posing substantial deployment challenges. Although distilled Small Language Models (SLMs) significantly enhance efficiency, their performance suffers as they fail to follow LLMs' reasoning paths. Luckily, we reveal that only a small fraction of tokens genuinely diverge reasoning paths between LLMs and SLMs. Most generated tokens are either identical or exhibit neutral differences, such as minor variations in abbreviations or expressions. Leveraging this insight, we introduce **Roads to Rome (R2R)**, a neural token router that selectively utilizes LLMs only for these critical, path-divergent tokens, while leaving the majority of token generation to the SLM. We also develop an automatic data generation pipeline that identifies divergent tokens and generates token-level routing labels to train the lightweight router. We apply R2R to combine R1-1.5B and R1-32B models from the DeepSeek family, and evaluate on challenging math, coding, and QA benchmarks. With an average activated parameter size of 5.6B, R2R surpasses the average accuracy of R1-7B by 1.6×, outperforming even the R1-14B model. Compared to R1-32B, it delivers a 2.8× wall-clock speedup with comparable performance, advancing the Pareto frontier of test-time scaling efficiency.

Despite rapid progress on AI benchmarks, the real-world meaning of benchmark performance remains unclear. To quantify the capabilities of AI systems in terms of human capabilities, we propose a new metric: 50%-task-completion time horizon. This is the time humans typically take to complete tasks that AI models can complete with 50% success rate. We first timed humans with relevant domain expertise on a combination of RE-Bench, HCAST, and 66 novel shorter tasks. On these tasks, current frontier AI models such as o3 have a 50% time horizon of around 110 minutes. Furthermore, frontier AI time horizon has been doubling approximately every seven months since 2019, though the trend may have accelerated since 2024. The increase in AI models’ time horizons seems to be primarily driven by greater reliability and ability to adapt to mistakes, combined with better logical reasoning and tool use capabilities. We discuss the limitations of our results—including their degree of external validity—and the implications of increased autonomy for dangerous capabilities. If these results generalize to real-world software tasks, extrapolation of this trend predicts that within 5 years, AI systems will be capable of automating many software tasks that currently take humans a month.

Despite recent advances in Vision-Language Models (VLMs), long-video understanding remains a challenging problem. Although state-of-the-art long-context VLMs can process around 1000 input frames, they still struggle to effectively leverage this sequence length, and succumb to irrelevant distractors within the context window. We present Dynamic Context Aggregation, an inference strategy for video question-answering that curates the model's input context. We use the VLM itself to iteratively identify and extract the most relevant frames from the video, which are then used for answering. We demonstrate how leveraging more computation at inference-time to select the most relevant context leads to improvements in accuracy, in agreement with recent work on inference-time scaling of LLMs. Moreover, we achieve state-of-the-art results on 4 diverse video question-answering datasets, showing consistent improvements with 3 different VLMs. In particular, our method shines on longer videos which would not otherwise fit in the model's context window: On longer videos of more than 1 hour on LVBench, our approach using a context window of 32K outperforms the same VLM using standard inference with a 700K context window by 2.8 points.

In-Context Learning (ICL) is a technique where large language models (LLMs) leverage multiple demonstrations (i.e., examples) to perform tasks. With the recent expansion of LLM context windows, many-shot ICL (generally with more than 50 demonstrations) can lead to significant performance improvements on a variety of language tasks such as text classification and question answering. Nevertheless, ICL faces the issue of demonstration order instability (ICL-DOI), which means that performance varies significantly depending on the order of demonstrations. Moreover, ICL-DOI persists in many-shot ICL, validated by our thorough experimental investigation. Current strategies for handling ICL-DOI are not applicable to many-shot ICL due to two critical challenges: (1) Most existing methods assess demonstration order quality by first prompting the LLM, then using heuristic metrics based on the LLM's predictions. In the many-shot scenarios, these metrics without theoretical grounding become unreliable, where the LLMs struggle to effectively utilize information from long input contexts, making order distinctions less clear. The requirement to examine all orders for the large number of demonstrations is computationally infeasible due to the super-exponential complexity of the order space in many-shot ICL. To tackle the first challenge, we design a demonstration order evaluation metric based on information theory for measuring order quality, which effectively quantifies the usable information gain of a given demonstration order. To address the second challenge, we propose a hierarchical demonstration order optimization method named \texttt{HIDO} that enables a more refined exploration of the order space, achieving high ICL performance without the need to evaluate all possible orders. Extensive experiments on multiple LLMs and real-world datasets demonstrate that our \texttt{HIDO} method consistently and efficiently outperforms other baselines. Our code project can be found at https://github.com/YinhanHe123/HIDO/.

Neural PDE solvers offer a powerful tool for modeling complex dynamical systems, but often struggle with error accumulation over long time horizons and maintaining stability and physical consistency. We introduce a multiscale implicit neural emulator that enhances long-term prediction accuracy by conditioning on a hierarchy of lower-dimensional future state representations. Drawing inspiration from the stability properties of numerical implicit time-stepping methods, our approach leverages predictions several steps ahead in time at increasing compression rates for next-timestep refinements. By actively adjusting the temporal downsampling ratios, our design enables the model to capture dynamics across multiple granularities and enforce long-range temporal coherence. Experiments on turbulent fluid dynamics show that our method achieves high short-term accuracy and produces long-term stable forecasts, significantly outperforming autoregressive baselines while adding minimal computational overhead.

Transformer-based Large Language Models (LLMs) have become increasingly important. However, scaling LLMs to longer contexts incurs slow inference speed and high GPU memory consumption for caching key-value (KV) vectors. This paper presents RetrievalAttention, a training-free approach to both accelerate the decoding phase and reduce GPU memory consumption by pre-building KV vector indexes for fixed contexts and maintaining them in CPU memory for efficient retrieval. Unlike conventional KV cache methods, RetrievalAttention integrate approximate nearest neighbor search (ANNS) indexes into attention computation. We observe that off-the-shelf ANNS techniques often fail due to the out-of-distribution (OOD) nature of query and key vectors in attention mechanisms. RetrievalAttention overcomes this with an attention-aware vector index. Our evaluation shows RetrievalAttention achieves near full attention accuracy while accessing only 1-3\% of the data, significantly reducing inference costs. Remarkably, RetrievalAttention enables LLMs with 8B parameters to handle 128K tokens on a single NVIDIA RTX4090 (24GB), achieving a decoding speed of 0.107 seconds per token.

Large language models (LLMs) are equipped with increasingly extended context windows recently, yet their long context understanding capabilities over long dependency tasks remain fundamentally limited and underexplored. This gap is especially significant in many real-world long-context applications that were rarely benchmarked. In this paper, we introduce $\textbf{LooGLE v2}$, a novel benchmark designed to evaluate LLMs' long context ability in real-world applications and scenarios. Our benchmark consists of automatically collected real-world long texts, ranging from 16k to 2M tokens, encompassing domains in law, finance, game and code. Accordingly, we delicately design 10 types of domain-specific long-dependency tasks and generate 1,934 QA instances with various diversity and complexity in a scalable data curation pipeline for further practical needs. We conduct a comprehensive assessment of 6 locally deployed and 4 API-based LLMs. The evaluation results show that even the best-performing model achieves only a 59.2\% overall score on our benchmark. Despite the extensive context windows, popular LLMs are only capable of understanding a much shorter length of context than they claim to be, revealing significant limitations in their ability to handle real-world tasks with long dependencies and highlighting substantial room for model improvement in practical long-context understanding.

Jailbreak attacks against large language models (LLMs) aim to induce harmful behaviors in LLMs through carefully crafted adversarial prompts. To mitigate attacks, one way is to perform adversarial training (AT)-based alignment, i.e., training LLMs on some of the most adversarial prompts to help them learn how to behave safely under attacks. During AT, the length of adversarial prompts plays a critical role in the robustness of aligned LLMs. While long-length adversarial prompts during AT might lead to strong LLM robustness, their synthesis however is very resource-consuming, which may limit the application of LLM AT. This paper focuses on adversarial suffix jailbreak attacks and unveils that to defend against a jailbreak attack with an adversarial suffix of length $\Theta(M)$, it is enough to align LLMs on prompts with adversarial suffixes of length $\Theta(\sqrt{M})$. Theoretically, we analyze the adversarial in-context learning of linear transformers on linear regression tasks and prove a robust generalization bound for trained transformers. The bound depends on the term $\Theta(\sqrt{M_{\text{test}}}/M_{\text{train}})$, where $M_{\text{train}}$ and $M_{\text{test}}$ are the numbers of adversarially perturbed in-context samples during training and testing. Empirically, we conduct AT on popular open-source LLMs and evaluate their robustness against jailbreak attacks of different adversarial suffix lengths. Results confirm a positive correlation between the attack success rate and the ratio of the square root of the adversarial suffix length during jailbreaking to the length during AT. Our findings show that it is practical to defend against "long-length" jailbreak attacks via efficient "short-length" AT. The code is available at https://github.com/fshp971/adv-icl.

Over more than a decade there has been an extensive research effort on how to effectively utilize recurrent models and attention. While recurrent models aim to compress the data into a fixed-size memory (called hidden state), attention allows attending to the entire context window, capturing the direct dependencies of all tokens. This more accurate modeling of dependencies, however, comes with a quadratic cost, limiting the model to a fixed-length context. We present a neural long-term memory module that learns to memorize historical context and helps attention to attend to the current context while utilizing long-past information. We show that this neural memory has the advantage of fast parallelizable training. From a memory perspective, we argue that attention due to its limited context but accurate dependency modeling performs as a short-term memory, while neural memory due to its ability to memorize the data, acts as a long-term, more persistent, memory. Based on these two modules, we introduce a new family of architectures, called Titans, and present three variants to address how one can effectively incorporate memory into this architecture. Our experimental results on language modeling, common-sense reasoning, and time series tasks show that Titans are effective compared to baselines, while they can effectively scale to larger context window in needle-in-haystack tasks.