长上下文

The key-value (KV) cache plays a vital role in accelerating autoregressive inference for large language models (LLMs). However, its linear memory growth with sequence length poses significant memory bottlenecks, especially in long-context scenarios. Quantization offers a promising solution for memory efficiency. While existing methods typically apply channel-wise quantization to the key cache and token-wise quantization to the value cache, they suffer from severe performance degradation under low-bit configurations. Our analysis reveals that quantization sensitivity varies across individual KV channels, presenting an opportunity for non-uniform bit allocation. Following this finding, we propose ChanMix, a mixed-precision quantization framework that supports channel-wise quantization on 2-bit setting with FP8 precision with a custom Triton kernel implementation. To improve low-bit quantization performance, we introduce a channel-aware bit reallocation strategy, which allocates bits across channel sensitivity. Through extensive evaluation, ChanMix demonstrates superior performance across the NIAH, RULER, and InfiniteBench benchmarks for the Llama, Mistral, and Qwen model families, achieving improvements of at least 5 absolute percentage points on RULER compared to all baseline methods. Additionally, ChanMix enables a 2.3× increase in batch size and supports a 1.5× longer context length during inference.

Long-context models (LCMs) have demonstrated great potential in processing long sequences, facilitating many real-world applications. The success of LCMs can be attributed to their ability to locate implicit critical information within the context for further prediction. However, recent research reveals that LCMs are often susceptible to contextual noise, i.e., irrelevant tokens, that can mislead model attention. In this paper, we conduct a fine-grained analysis of the context noise and propose an effective metric, the Integrated Gradient (IG) score, to detect and quantify the noise information within the context. Our findings reveal that even simple mitigation of detected context noise can substantially boost the model's attention on critical tokens and benefit subsequent predictions. Building on this insight, we propose Context Denoising Training (CDT), a straightforward yet effective training strategy that improves attention on critical tokens while reinforcing their influence on model predictions. Extensive experiments across four tasks, under both context window scaling and long-context alignment settings, demonstrate the superiority of CDT. Notably, when trained with CDT, an open-source 8B model can achieve performance (50.92) comparable to GPT-4o (51.00).

Learning to simulate complex physical systems from data has emerged as a promising way to overcome the limitations of traditional numerical solvers, which often require prohibitive computational costs for high-fidelity solutions. Recent Graph Neural Simulators (GNSs) accelerate simulations by learning dynamics on graph-structured data, yet often struggle to capture long-range interactions and suffer from error accumulation under autoregressive rollouts. To address these challenges, we propose Information-preserving Graph Neural Simulators (IGNS), a graph-based neural simulator built on the principles of Hamiltonian dynamics. This structure guarantees preservation of information across the graph, while extending to port-Hamiltonian systems allows the model to capture a broader class of dynamics, including non-conservative effects. IGNS further incorporates a warmup phase to initialize global context, geometric encoding to handle irregular meshes, and a multi-step training objective that facilitates PDE matching, where the trajectory produced by integrating the port-Hamiltonian core aligns with the ground-truth trajectory, thereby reducing rollout error. To evaluate these properties systematically, we introduce new benchmarks that target long-range dependencies and challenging external forcing scenarios. Across all tasks, IGNS consistently outperforms state-of-the-art GNSs, achieving higher accuracy and stability under challenging and complex dynamical systems. Our project page: https://thobotics.github.io/neural_pde_matching.

Existing key-value (KV) cache compression methods for large language models (LLMs) often rely on token eviction, which risks losing critical local information in both long prefilling and decoding scenarios. When extrapolating beyond the pretrained context length, their performance degrades sharply on long-context benchmarks. Motivated by the observation in the frequency domain that the context information is concentrated in the low-frequency components, we propose FreqKV, a parameter-free and architecture-agnostic approach. It iteratively compresses the increasing KV cache in the frequency domain, allowing models to process lengthy contexts efficiently. With minimal training at 8K length, FreqKV extends the context window of LLaMA-2-7B up to 256K tokens while maintaining stable perplexity. Extensive experiments on both prefilling and decoding stages demonstrate that FreqKV enables robust context window extension and consistently outperforms existing KV cache compression methods, highlighting its effectiveness for both understanding and generation in long contexts.

There has been work discussing the impact of long context on Language Model performance: some find that long irrelevant context could harm performance, while some experimentally summarize loss reduction by relevant long context as Scaling Laws. This calls for a more thorough understanding on how long context impacts Language Modeling. In this work, we (1) propose to use Intrinsic Entropy for explaining the impact of context length on language modeling; and (2) conduct experiments on natural language and synthetic data, validating our proposed theoretical assumptions and deductions. Our theoretical framework can provide practical insights such as establishing that training dataset size dictates an optimal context length and bounds context length scaling for certain cases. We hope our work may inspire new long context Language Models, as well as future work studying Physics for Language Models.

Long-context models are essential for many applications but face inefficiencies in loading large KV caches during decoding. Prior methods enforce fixed token budgets for sparse attention, assuming a set number of tokens can approximate full attention. However, these methods overlook variations in the importance of attention across heads, layers, and contexts. To address these limitations, we propose Tactic, a sparsity-adaptive and calibration-free sparse attention mechanism that dynamically selects tokens based on their cumulative attention scores rather than a fixed token budget. By setting a target fraction of total attention scores, Tactic ensures that token selection naturally adapts to variations in attention sparsity. To efficiently approximate this selection, Tactic leverages clustering-based sorting and distribution fitting, allowing it to accurately estimate token importance with minimal computational overhead. We show that Tactic outperforms existing sparse attention algorithms, achieving superior accuracy and up to 5.14x decode attention speedup. This improvement translates to an overall 1.51x end-to-end inference speedup, making Tactic a practical and effective solution for long-context LLM inference in accuracy-sensitive applications.

Long-horizon interactions between users and LLM-based assistants necessitates effective memory management, yet current approaches face challenges in training and evaluation of memory. Existing memory benchmarks rely on static, off-policy data as context, limiting evaluation reliability and scalability. To address these gaps, we introduce AMemGym, an interactive environment enabling on-policy evaluation and optimization for memory-driven personalization. AMemGym employs structured data sampling to predefine user profiles, state-dependent questions, and state evolution trajectories, enabling cost-effective generation of high-quality, evaluation-aligned interactions. LLM-simulated users expose latent states through role-play while maintaining structured state consistency. Comprehensive metrics based on structured data guide both assessment and optimization of assistants. Extensive experiments reveal performance gaps in existing memory systems (e.g., RAG, long-context LLMs, and agentic memory) and corresponding reasons. AMemGym not only enables effective selection among competing approaches but also can potentially drive the self-evolution of memory management strategies. By bridging structured state evolution with free-form interactions, our framework provides a scalable, diagnostically rich environment for advancing memory capabilities in conversational agents.

The quadratic complexity of self-attention mechanism presents a significant impediment to applying Transformer models to long sequences. This work explores computational principles derived from astrocytes—glial cells critical for biological memory and synaptic modulation—as a complementary approach to conventional architectural modifications for efficient self-attention. We introduce the Recurrent Memory Augmented Astromorphic Transformer (RMAAT), an architecture integrating abstracted astrocyte functionalities. RMAAT employs a recurrent, segment-based processing strategy where persistent memory tokens propagate contextual information. An adaptive compression mechanism, governed by a novel retention factor derived from simulated astrocyte long-term plasticity (LTP), modulates these tokens. Attention within segments utilizes an efficient, linear-complexity mechanism inspired by astrocyte short-term plasticity (STP). Training is performed using Astrocytic Memory Replay Backpropagation (AMRB), a novel algorithm designed for memory efficiency in recurrent networks. Evaluations on the Long Range Arena (LRA) benchmark demonstrate RMAAT's competitive accuracy and substantial improvements in computational and memory efficiency, indicating the potential of incorporating astrocyte-inspired dynamics into scalable sequence models.

Long-term time series forecasting (LTSF) plays a crucial role in fields such as energy management, finance, and traffic prediction. Transformer-based models have adopted patch-based strategies to capture long-range dependencies, but accurately modeling shape similarities across patches and variables remains challenging due to scale differences. To address this, we introduce patch-mean decoupling (PMD), which separates the trend and residual shape information by subtracting the mean of each patch, preserving the original structure and ensuring that the attention mechanism captures true shape similarities. Futhermore, to more effectively model long-range dependencies and capture cross-variable relationships, we propose Trend Restoration Attention (TRA) and Proximal Variable Attention (PVA). The former module reintegrates the decoupled trend from PMD while calculating attention output. And the latter focuses cross-variable attention on the most relevant, recent time segments to avoid overfitting on outdated correlations. Combining these components, we propose PMDformer, a model designed to effectively capture shape similarity in long-term forecasting scenarios. Extensive experiments indicate that PMDformer outperforms existing state-of-the-art methods in stability and accuracy across multiple LTSF benchmarks.

Long-context reasoning requires accurately identifying relevant information in extensive, noisy input contexts. In this work, we propose PERK (Parameter Efficient Reasoning over Knowledge), a scalable approach for learning to encode long contexts using gradient updates at test time. Specifically, PERK employs two nested optimization loops in a meta-training phase. The inner loop rapidly encodes contexts into a low-rank adapter (LoRA) that serves as a parameter-efficient memory module for the base model. Concurrently, the outer loop learns to use the updated adapter to accurately recall and reason over relevant information from the encoded long context. Our evaluations on several long-context reasoning tasks show that PERK significantly outperforms the standard long-context finetuning, achieving average absolute performance gains of up to 20% for Qwen-2.5 (0.5B & 7B) on synthetic and real-world long-context reasoning. PERK also maintains its advantages across model scales and families. Compared to specialized long-context LLMs, PERK matches or surpasses their performance. Finally, our analyses show PERK is more robust to reasoning complexity, length extrapolation, and the positions of relevant information in contexts.

Recent advances in deep-research agents have shown promise for autonomous knowledge construction through dynamic reasoning over external sources. However, existing approaches rely on a mono-contextual paradigm that accumulates all information in a single, expanding context window, leading to context suffocation and noise contamination that limit their effectiveness on long-horizon tasks. We introduce IterResearch, a novel iterative deep-research paradigm that reformulates long-horizon research as a Markov Decision Process with strategic workspace reconstruction. By maintaining an evolving report as memory and periodically synthesizing insights, our approach preserves consistent reasoning capacity across arbitrary exploration depths. We further develop Efficiency-Aware Policy Optimization (EAPO), a reinforcement learning framework that incentivizes efficient exploration through geometric reward discounting and enables stable distributed training via adaptive downsampling. Extensive experiments demonstrate that IterResearch achieves substantial improvements over existing open-source agents with average +14.5pp across six benchmarks and narrows the gap with frontier proprietary systems. Remarkably, our paradigm exhibits unprecedented interaction scaling, extending to 2048 interactions with dramatic performance gains (from 3.5\% to 42.5\%), and serves as an effective prompting strategy, improving frontier models by up to 19.2pp over ReAct on long-horizon tasks. These findings position IterResearch as a versatile solution for long-horizon reasoning, effective both as a trained agent and as a prompting paradigm for frontier models.

Memory and computation remain core bottlenecks in long-horizon LLM inference due to the quadratic cost of self-attention and the ever-growing key-value (KV) cache. Existing strategies for memory-bounded inference, such as quantization, offloading, or heuristic KV eviction, either incur high orchestration costs or rely on unreliable attention-based proxies of importance. We propose TRIM-KV, a novel approach that learns each token’s intrinsic importance at creation time via a lightweight retention gate. Each gate predicts a scalar retention score that decays over time, reflecting the long-term utility of the token for a specific layer and head. Tokens with low scores are evicted when the memory budget is exceeded, ensuring that the cache always contains the most critical tokens. TRIM-KV is trained efficiently through distillation from a frozen LLM combined with a capacity loss, requiring only gate fine-tuning and adding negligible inference overhead. Across mathematical reasoning (GSM8K, MATH-500, AIME24), procedural generation (LongProc), and conversational long-memory benchmarks (LongMemEval), TRIM-KV consistently outperforms strong eviction and learnable retrieval baselines, especially in low-memory regimes. Remarkably, it even surpasses full-cache models in some settings, showing that selective retention can serve as a form of regularization, suppressing noise from uninformative tokens. Qualitative analyses further reveal that learned retention scores align with human intuition, naturally recovering heuristics such as sink tokens, sliding windows, and gist compression without explicit design. Beyond efficiency, retention scores provide insights into layer- and head-specific roles, suggesting a new path toward LLM interpretability.

This paper introduces **Story-Iter**, a new training-free iterative paradigm to enhance long-story generation. Unlike existing methods that rely on fixed reference images to construct a complete story, our approach features a novel external **iterative paradigm**, extending beyond the internal iterative denoising steps of diffusion models, to continuously refine each generated image by incorporating all reference images from the previous round. To achieve this, we propose a plug-and-play, training-free **g**lobal **r**eference **c**ross-**a**ttention (**GRCA**) module, modeling all reference frames with global embeddings, ensuring semantic consistency in long sequences. By progressively incorporating holistic visual context and text constraints, our iterative paradigm enables precise generation with fine-grained interactions, optimizing the story visualization step-by-step. Extensive experiments in the official story visualization dataset and our long story benchmark demonstrate that Story-Iter's state-of-the-art performance in long-story visualization (up to 100 frames) excels in both semantic consistency and fine-grained interactions.

This paper studies the expressive power of linear attention layers for in-context learning (ICL) of linear dynamical systems (LDS). We consider training on sequences of inexact observations produced by noise-corrupted LDSs, with all perturbations being Gaussian; importantly, we study the non-i.i.d. setting as it is closer to real-world scenarios. We provide the optimal weight construction for a single linear-attention layer and show its equivalence to one step of Gradient Descent relative to an autoregression objective of window size one. Guided by experiments, we uncover a relation to the Preconditioned Conjugate Gradient method for larger window sizes. We back our findings with numerical evidence. These results add to the existing understanding of transformers' expressivity as in-context learners, and offer plausible hypotheses for experimental observations whereby they compete with Kalman filters --- the optimal model-dependent learners for this setting.

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Large Language Models (LLMs) by optimizing them against factual outcomes. However, this paradigm falters in long-context scenarios, as its reliance on internal parametric knowledge is ill-suited for tasks requiring contextual grounding—the ability to find and reason over externally provided information. We identify a key reason for this failure: a reward based solely on the final answer is too sparse to effectively guide the model for identifying relevant evidence. We formally prove that the outcome-only reward leads to exponentially vanishing gradients for the context grounding process, rendering learning intractable. To overcome this bottleneck, we introduce LongRLVR to augment the sparse answer reward with a dense and verifiable context reward. This auxiliary signal directly incentivizes the model for selecting the correct grounding information, providing a robust learning gradient that solves the underlying optimization challenge. We validate our method on challenging long-context benchmarks using Qwen and LLaMA models. LongRLVR consistently and significantly outperforms the standard RLVR across all models and benchmarks, e.g., boosting a 14B model's scores on RULER-QA from 73.17 to 88.90 and on LongBench v2 from 39.8 to 46.5. Our work demonstrates that explicitly rewarding the grounding process is a critical and effective strategy for unlocking the full reasoning potential of LLMs in long-context applications.

Reasoning over long contexts is essential for large language models. While reinforcement learning (RL) enhances short-context reasoning by inducing "Aha" moments in chain-of-thought, the advanced thinking patterns required for long-context reasoning remain largely unexplored, and high-difficulty RL data are scarce. In this paper, we introduce LoongRL, a data-driven RL method for advanced long-context reasoning. Central to LoongRL is KeyChain, a synthesis approach that transforms short multi-hop QA into high-difficulty long-context tasks by inserting UUID chains that hide the true question among large collections of distracting documents. Solving these tasks requires the model to trace the correct chain step-by-step, identify the true question, retrieve relevant facts and reason over them to answer correctly. RL training on KeyChain data induces an emergent plan–retrieve–reason–recheck reasoning pattern that generalizes far beyond training length. Models trained at 16K effectively solve 128K tasks without prohibitive full-length RL rollout costs. On Qwen2.5-7B and 14B, LoongRL substantially improves long-context multi-hop QA accuracy by +23.5% and +21.1% absolute gains. The resulting LoongRL-14B reaches a score of 74.2, rivaling much larger frontier models such as o3-mini (74.5) and DeepSeek-R1 (74.9). It also improves long-context retrieval, passes all 128K needle-in-a-haystack stress tests, and preserves short-context reasoning capabilities.

The limited context window of contemporary large language models (LLMs) hinders broader application. In this work, we present SharedLLM, a novel approach grounded in the design philosophy of multi-grained context compression and query-aware information retrieval. SharedLLM is composed of two short-context LLMs: a lower moel (compressor) and an upper model (decoder). The lower model compresses context information, while the upper model processes compressed, context information from the lower model and performs context-aware modeling. Information transfer between the compressor and decoder occurs only at the lowest layers to reduce redundant computation. Based on this architecture, we introduce a specialized tree-style data structure to efficiently encode, store and retrieve multi-grained contextual information from text chunks. This entire process, wherein the sender and receiver are derived from the same LLM layer, is referred to as self-injection. In our evaluation on long-context modeling and understanding tasks, SharedLLM achieves superior or comparable results to several strong baselines, striking an effective balance between efficiency and performance. Meanwhile, with the aforementioned design choices, SharedLLM can greatly reduce memory consumption, and demonstrates substantial speed-ups over other advanced baselines. The core code of our implementation along with training and evaluation is available in appendix and supplementary.

While transformer-based language models have driven the AI revolution thus far, their computational complexity has spurred growing interest in viable alternatives, such as structured state space sequence models (SSMs) and Selective SSMs. Among these, Mamba (S6) and its variant Mamba-2 have shown remarkable inference speed-ups over transformers while achieving comparable or superior performance on complex language modeling tasks. However, despite these architectural innovations and empirical successes, the fundamental learning capabilities of Mamba remain poorly understood. In this paper, we address this gap by studying in-context learning (ICL) on Markov chains and uncovering an interesting phenomenon: even a single-layer Mamba efficiently learns the in-context Laplacian smoothing estimator, which is both Bayes and minimax optimal. To explain this, we theoretically characterize the representation capacity of Mamba and reveal the fundamental role of convolution in enabling it to represent the optimal Laplacian smoothing. These theoretical insights align strongly with empirical results and, to the best of our knowledge, represent the first formal connection between Mamba and optimal statistical estimators. Finally, we outline promising research directions inspired by these findings.

Reinforcement learning (RL) is a human-designed framework for solving sequential decision-making problems. In this work, we demonstrate that, surprisingly, RL emerges in LLMs at inference time – a phenomenon known as in-context RL (ICRL). To reveal this capability, we introduce a simple multi-round prompting framework, called ICRL prompting. The goal of ICRL prompting is to guide LLMs to perform reinforcement learning for self-improvement on a given task. After the LLM generates a response at the current round, we give numerical scalar feedback on the response, called the rewards. At the next round, we prompt the LLM again with the same task and a context consisting of all previous responses and rewards. We observe that the quality of the LLM's response increases as the context grows. In other words, the LLM is able to maximize the scalar reward signal at inference time, just like an RL algorithm. We evaluate ICRL prompting on Game of 24, creative writing, ScienceWorld, and Olympiad-level math competitions (AIME and HMMT), demonstrating significant improvements over baselines such as Self-Refine and Reflexion. Surprisingly, in some experiments, the reward signals are generated by the LLM itself, yet performance improvements are still observed from ICRL prompting, offering a new paradigm for test-time scaling.

Understanding how long-term stimulation reshapes neural circuits requires uncovering the rules of brain plasticity. While short-term synaptic modifications have been extensively characterized, the principles that drive circuit-level reorganization across hours to weeks remain unknown. Here, we formalize these principles as a latent dynamical law that governs how recurrent connectivity evolves under repeated interventions. To capture this law, we introduce the Stimulus-Evoked Evolution Recurrent dynamics (STEER) framework, a dual-timescale model that disentangles fast neural activity from slow plastic changes. STEER represents plasticity as low-dimensional latent coefficients evolving under a learnable recurrence, enabling testable inference of plasticity rules rather than absorbing them into black-box parameters. We validate STEER with four benchmarks: synthetic Lorenz systems with controlled parameter shifts, BCM-based networks with biologically grounded plasticity, a task learning setting with adaptively optimized external stimulation and longitudinal recordings from Parkinsonian rats receiving closed-loop DBS. Our results demonstrate that STEER recovers interpretable update equations, predicts network adaptation under unseen stimulation schedules, and supports the design of improved intervention protocols. By elevating long-term plasticity from a hidden confound to an identifiable dynamical object, STEER provides a data-driven foundation for both mechanistic insight and principled optimization of brain stimulation.