Efficient Long-Context Inference

Context engineering decides what information a model carries forward, and current designs meter it in tokens: compressing the past into a bounded recurrent state, keeping a key-value entry for every token, or imposing a fixed budget through a window or eviction rule. All three make the token the unit of memory even when the stream is redundant and the task depends on the distinct information it carries. Building on a companion mechanism paper that opens a cache slot only when an incoming key is novel, so memory scales with the number of distinct items rather than tokens, we develop that allocate-on-novelty cache as a working-memory component and organize context by how a task depends on the past: recall-carried information belongs in a content-addressed novelty cache, summary-carried information in a recurrent state, and locality-carried information in a recency window. The claim is empirical and bounded. On a matched character-level control, novelty-gated attention reaches full-attention performance while attending to about half the tokens, and coupling the cache with a state-space summary matches full-attention coupling at that reduced cost; the advantage grows as context lengthens, while a sliding window is preferable on short, locality-dominated spans. On next-code prediction over synthetic Medicare claims the coupled component leads full attention and every fixed-budget eviction policy at a thousand-event horizon, whereas cost forecasting over the same stream is summary-carried and the cache is neutral. The retained memory is an inspectable table of templates, codes, drugs, or places rather than an opaque state. The experiments are small-scale and use only public data; they establish the primitive that context can scale with distinct information rather than tokens, in a working memory that is content-addressable and auditable.

Long-context processing has become increasingly important for large language models (LLMs), but simply extending the context window does not guarantee effective utilization of long inputs. As input length grows, accuracy often degrades, indicating that models still struggle to identify and use the evidence most relevant to a question. A promising way to improve long-context utilization is test-time training (TTT), which treats the test context as a training example for instance-specific parameter adaptation. However, applying TTT to the entire long context is prohibitively expensive, while adapting on randomly sampled spans introduces severe noise. Because most spans in a long context are irrelevant to the specific question, training on them may even degrade the base model's performance. Our preliminary study shows that TTT is highly sensitive to training-span quality: on LongBench-v2, TTT on randomly sampled spans hurts performance, whereas TTT on oracle spans substantially improves it. Motivated by this, we propose a simple method, Self-Guided TTT (S-TTT): before adaptation, the model identifies the evidence spans it should learn from, and the standard language-modeling training objective is applied only to those selected spans. On two challenging long-context reasoning benchmarks, LongBench-v2 and LongBench-Pro, S-TTT improves accuracy for both Qwen3-4B-Thinking-2507 and Llama-3.1-8B-Instruct, achieving up to a 15% relative improvement.

Large language models, and the agents built on them, spend an ever-growing share of their compute and memory on remembering: caching attention keys and values, carrying long prompts, maintaining recurrent state, and storing what happened in previous turns and sessions. Because none of this memory is free, four largely separate research communities have each learned to compact it. They evict or quantize the KV cache, prune or distill prompts, bound architectural state, and consolidate agent memory. We argue that these are instances of one problem: a rate--distortion decision about what context-derived information to retain versus discard, at what fidelity, under a resource budget, so as to preserve downstream task utility. We make this lens precise with a single compaction objective and a layer-agnostic lower bound, use it to build a seven-axis taxonomy that classifies methods from across the stack uniformly, and use it to transfer mechanisms between layers that have never been connected, from serving-stack KV management to agent long-term memory. Two patterns hold across the survey. At every layer the signal that decides what to keep is attention magnitude or recency, and it fails in the same way everywhere, by discarding, before the query is known and with no way to undo it, information the query later needs. And while compression is measured carefully on single-turn long context, the repeated compaction that agents actually perform is almost never measured, and no benchmark holds one budget axis across all the layers at once. We turn both observations into a benchmark proposal, a small reference experiment, and a set of compaction-aware design principles, and we map the open problems.

The quadratic cost of causal self-attention severely bottlenecks long-context transformer inference. While numerous post hoc linearization pipelines exist, it is difficult to identify which components preserve model quality. This work isolates the effect of state update design in a strict frozen-backbone regime. We show that softmax relies on key-dependent, rank-1 orthogonal projections, elucidating why delta-style networks outperform purely gated accumulation. We identify a potential source of approximation errors and introduce structural interventions, specifically sink tokens, short convolutions, and fixed-budget cache routing, which reduces the remaining gap. We scale this linearization approach across LLaMA and Qwen models up to 32B parameters, outperforming prior post hoc baselines on MMLU and matching the long-context retrieval of complex adaptive-caching frameworks.

Speculative decoding accelerates LLM inference by drafting multiple tokens and verifying them in parallel. Block-parallel drafters such as DFlash further improve drafting efficiency by predicting an entire block in one pass, but their position-wise predictions lack explicit intra-block causal conditioning. Recent methods such as Domino and DSpark attempt to introduce such causality into block-parallel drafting, but they require training the draft model from scratch, which limits their flexibility and increases training cost. We propose DeLS-Spec, a decoupled long-short context speculative decoding method. DeLS-Spec treats the fixed DFlash model as a long-context expert and introduces a lightweight local head as a short-context expert. The local head can be trained independently with a standard next-token prediction objective, without joint training with the target model or the DFlash backbone, leading to extremely low training cost. At inference time, DeLS-Spec combines long-context and short-context logits, and the local head is not tied to a specific DFlash checkpoint, making the method more modular and flexible. Experiments on Qwen3 models show that DeLS-Spec consistently improves speedup and average acceptance length over DFlash across math, code, and dialogue benchmarks.

Linear attention models allow a fixed state size and a fixed amount of compute per token. However, due to their limited state size, linear attention models fall behind in long-context recall compared to softmax-attention-based transformer architectures. Increasing the state size of linear attention improves recall performance but at the cost of higher FLOPs. In this work, we introduce Sparse Delta Memory (SDM), an architecture that scales the hidden state of gated linear RNNs to orders of magnitude higher capacity using a sparse addressing scheme. SDM extends the Gated DeltaNet architecture by replacing the dense key-value outer product with sparse reads and writes to a large explicit memory. We show that, under an isoFLOP constraint and with an identical number of parameters, a higher state memory capacity significantly improves performance on in-context learning and long-context retrieval tasks. Moreover, by learning the initial state of the SDM memory and therefore using it as a parametric memory, we show that the model further improves on a wide range of common-knowledge and reasoning tasks.

Modern LLMs are increasingly deployed in long-context applications such as retrieval-augmented generation, repository-level coding, and agentic workflows whose accumulated reasoning and tool traces routinely push the input an order of magnitude past the pretraining window, making zero-shot context extension the dominant deployment path for open-weight checkpoints. The dominant zero-shot methods (YaRN, Self-Extend, DCA) fix a single rescaling factor up front, so an aggressive factor sacrifices short-context fidelity while a conservative one breaks down at long contexts; recent length-aware variants adapt the mapping, but with a fitted or distance-dependent schedule. We propose Jet-Long, a tuning-free zero-shot method that pairs a local RoPE-faithful window with a long-range window whose rescaling factor adapts dynamically to the current sequence length via a parameter-free analytic schedule, recovering the base model exactly at short inputs while extrapolating cleanly at long ones. An inclusion-exclusion attention merge and an on-the-fly RoPE correction rotation make the bifocal construction essentially free at inference; fused into a single CuTe kernel, long-context prefill reaches up to $1.39\times$ FA2 throughput on H100 (approaching the Hopper-only FA4), and single-batch generation incurs $\le 4\%$ overhead at every length. On Qwen3-1.7B/4B/8B up to 128K context, Jet-Long leads RULER by $+4.79$/$+2.18$/$+2.03$ pp over the strongest baseline at 1.7B/4B/8B, achieves the best overall accuracy on HELMET-RAG (a benchmark identified by HELMET as the most efficient predictor of downstream long-context performance) and attains the lowest PG-19 perplexity. Jet-Long also generalizes to hybrid attention architectures such as Jet-Nemotron for further long-context improvement without retraining, and remains hyperparameter-resilient for ease of deployment.

Long-context language model inference is increasingly limited by the memory bandwidth and capacity required to store key-value caches, yet existing compression methods often apply uniform budgets across layers or tokens and degrade retrieval when lexical cues and semantic states require different preservation. We introduce DepthWeave-KV, a token-adaptive cache compression method that factorizes key and value states across neighboring transformer layers using shared low-rank channel bases while retaining lightweight token-specific residuals where attention behavior is sensitive. DepthWeave-KV combines cross-depth residual factorization with a token-conditional depth router that allocates higher reconstruction rank to instruction-bearing and retrieval-critical tokens, and uses calibration-free online error tracking from attention-output probes to adapt compression during generation without retraining the base model. A fused CUDA implementation jointly performs basis lookup, residual dequantization, and attention projection to reduce decode-time memory traffic. Across LongBench, Needle-in-a-Haystack, L-Eval, and long-form QA and summarization benchmarks, DepthWeave-KV achieves near-full-cache task quality with substantially lower memory use, improving average score and retrieval accuracy over prior compressed caches while reaching 8.3x KV memory reduction and 72.8 tokens per second at 64K context.

Synthesizing long-context supervised fine-tuning (SFT) data is a scalable way to enhance the long-context understanding of large language models (LLMs), yet existing approaches share three limitations: narrow task coverage, insufficient instruction difficulty, and a lack of faithfulness supervision. We propose \textbf{LongCrafter}, a structured synthesis framework that couples a hierarchical task taxonomy with an evidence-grounded pipeline. The taxonomy organizes long-context understanding into local/shallow and global/deep levels and yields 32 fine-grained task types that serve as a global generative prior. Guided by this taxonomy, LongCrafter constructs task-aligned long contexts, decomposes them into explicit evidence graphs that model cross-paragraph dependencies, and generates instruction--response pairs strictly grounded in the located evidence spans, ensuring both controllable difficulty and faithful, traceable reasoning. Models fine-tuned on LongCrafter data outperform all SFT baselines and even the official post-trained models on LongBench, LongBench~v2, and LooGLE across both Qwen2.5-7B and LLaMA-3.1-8B, with the largest gains on high-difficulty tasks. Further analysis shows that LongCrafter data is more diverse and better spread across difficulty levels, and that the trained models locate evidence robustly regardless of position, effectively mitigating the ``lost in the middle'' problem.

Contemporary language models are dominated by the transformer architecture, which leverages self-attention mechanisms to enable more efficient, parallelized training across a wide set of documents and corpora. This has allowed transformers to effectively model data across a wide range of modalities and contexts. However, transformers, along with their conventional counterparts such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs), often struggle to maintain efficiency when processing long contexts. We introduce ResonatorLM, a new mechanism that replaces attention with a physics-derived alternative. ResonatorLM treats token sequences as a single, driven one-dimensional latent field and replaces attention dot products with causal functions of damped resonators. We implement ResonatorLM on a traditional network architecture and test it on standard long-context modeling tasks. We find that in a small, 6M matched setting, training and prefill speedups increase with sequence length, decode speed reaches 6.47x compared to that of a standard, optimized transformer at 32K tokens, and accuracy reaches 61.31 percent (compared to 55.32 percent) on WikiText.

Long-horizon agentic LLMs are increasingly limited by finite context windows, as extended interaction trajectories can exceed the maximum context length before a task is completed. Context compaction offers a natural solution by summarizing previous interaction states and continuing the rollout under a compressed context, but incorporating compaction into reinforcement learning remains underexplored. We propose CompactionRL, a reinforcement learning strategy to train long-horizon agentic LLMs with context compaction. Our approach jointly optimizes task execution and summary generation with token-level loss normalization and cross-trajectory generalized advantage estimation. This design enables the LLM agents to learn from compacted long-horizon trajectories. We train CompactionRL on top of open models and observe consistent performance gains on agentic coding tasks. CompactionRL enables the open GLM-4.5-Air model (106B-A30B) to achieve Pass@1 scores of 66.8% on SWE-bench Verified and 24.5% on Terminal-Bench 2.0, with absolute gains of 7.0 and 3.1 points, respectively. Built upon GLM-4.7-Flash (30B-A3B), CompactionRL improves Pass@1 by 5.5 and 6.8 points, reaching 56.0% on SWE-bench Verified and 20.2% on Terminal-Bench 2.0, respectively. CompactionRL is thus deployed in the RL pipeline for training the open GLM-5.2 model (750B-A40B).

We present Nemotron-Labs-3-Puzzle-75B-A9B, a compressed variant of Nemotron-3-Super optimized for interactive deployment. We designed the model to maximize server throughput under high user throughput constraints. In interactive serving workloads on a single 8xB200 node, Puzzle-75B-A9B achieves approximately 2x higher server throughput than Nemotron-3-Super at matched user throughput constraints. In ultra-long-context deployment on a single H100 GPU, the compressed model increases 1M-token concurrency from 1 request to 8 requests. Puzzle-75B-A9B is constructed using a multi-stage pipeline that combines the Iterative Puzzle compression framework with knowledge distillation, reinforcement learning, quantization, and a Multi-Token Prediction head. The compression process jointly optimizes heterogeneous MoE pruning, active parameter budget, and Mamba pruning to improve inference efficiency while preserving model quality. We evaluate Puzzle-75B-A9B on a broad suite of reasoning, coding, multilingual, long-context, and agentic benchmarks. Despite substantial compression, the model retains strong downstream accuracy relative to the parent model across a wide range of tasks. These results demonstrate that large hybrid MoE models can be substantially optimized for deployment efficiency while maintaining strong downstream capability. Our model is publicly available on Hugging Face.

Portuguese remains underrepresented in text embedding evaluation, despite being one of the most widely spoken languages in the world. As a result, embedding models are often selected based on English or multilingual metrics, while their effectiveness in Portuguese remains unclear. We present MTEB-PT, a Portuguese benchmark constructed from a subset of MMTEB, comprising 14 existing datasets across Semantic Textual Similarity (STS), classification, retrieval, and reranking. We use this benchmark to evaluate 17 open- and closed-source embedding models under a unified protocol. Our results show that Portuguese performance is strongly task-dependent: multilingual rankings do not reliably predict Portuguese-specific performance across task families, no single model dominates all settings, and models with stronger long-context capacity are particularly advantageous on longer-input tasks such as retrieval and reranking. The benchmark also shows that language-specific fine-tuning still improves model performance in Portuguese, especially on task types that match the adaptation data most closely. To examine this effect, we fine-tune three representative backbone models with Portuguese contrastive supervision and Matryoshka Representation Learning (MRL). These benchmark-informed baselines yield their strongest gains on STS, consistent with the predominantly symmetric supervision used during training, while also improving retrieval and remaining competitive under dimensional truncation. We release the MTEB-PT benchmark, the fine-tuned models, and the training and evaluation code.

Block-sparse attention scales long-context language models by replacing the O(N^2) softmax with a per-query top-k selection over key blocks. This cutoff is myopic: when the k-th and (k+1)-th blocks are nearly tied in score, the selector commits without spending extra budget, and a dropped block carrying answer evidence is unrecoverable downstream. We propose a value-of-information router that measures, for each query, how decisively the top-k cut was made, and doubles the kept set for the queries where that gap is smallest; the rule is backbone-agnostic and stacks with existing block-scoring methods such as Quest. On LongBench-v2 medium at n=215 (the entire dataset subset), router-on-Quest reaches paired recall 0.75 vs. top-k 0.47 -- +28 pp over the SSA-style baseline (McNemar p<0.01) -- and lands within 2 pp of dense on RULER NIAH multikey at the same context. The lift reproduces on four models from three architectures (Qwen2.5, Mistral-Nemo, Qwen3.6). At 128K, the router preserves 0.81 and 0.89 of dense accuracy on Qwen2.5-7B-1M and Qwen3.6 (vs. SSA-style top-k at 0.09 on the former) while the fused selection-plus-kernel pipeline runs at 0.62x and 0.80x dense wall time.

Scaling modern large language models (LLMs) to long contexts is limited by the quadratic computation cost, and poor length extrapolation of dense attention. Chunk-wise sparse attention offers a promising alternative, but all existing methods fall short of full attention because of their inaccurate chunk selection. We propose Hierarchical Landmark Sparse (HiLS) Attention, a chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling (LM) loss. HiLS factorizes attention hierarchically: each query performs attention independently with each retrieved chunk to extract chunk-specific information, and the resulting outputs are fused according to chunk retrieval scores. By incorporating retrieval scores into the forward attention computation, HiLS optimizes them directly with the LM loss, enabling end-to-end retrieval learning and native sparse training. Experimental results show that HiLS-Attention achieves performance comparable to, and in some cases better than, full attention at in-domain context lengths. Meanwhile, HiLS-Attention extrapolates more than $64\times$ the training context length with 90% retrieval accuracy, far beyond full attention. Moreover, existing full-attention models can be converted to HiLS-Attention with lightweight continued pretraining, preserving in-domain performance while acquiring ultra-long-context extrapolation. Together with its sparse KV access and computation, HiLS-Attention breaks the usual efficiency-performance trade-off, enabling long-context LLMs that are both more efficient and more effective on general long-context tasks than their full-attention counterparts.

High-throughput long-context generation is one of the central challenges for large language models. Generation is typically memory-bandwidth-bound rather than compute-bound: each decoding step must stream the accumulated key/value (KV) cache from memory, so bandwidth demand grows with context length while only one token is emitted. Two parallel approaches have therefore emerged: reducing memory access with efficient attention variants and linear-time mixers such as Mamba, or increasing parallel computation by generating blocks of tokens at once. However, technical challenges arise when combining these two ideas. Earlier hybrid diffusion models such as DiffuMamba use bidirectional Mamba mixing, including a reverse-direction scan relative to causal generation. This reverse scan needs to scan the entire sequence, so its states are not prefix-only and cannot be precisely reused as a cache even when diffusion is performed block by block. We propose a BDLM Mamba--attention hybrid that addresses this challenge by restricting the reverse Mamba scan to the active denoising block, which enables exact caching across blocks. In an 87M-parameter DCLM sweep, BDLM Mamba-H achieves the best C4-en validation perplexity compared to BDLM attention and full-sequence baselines. At 350M parameters, it remains competitive with BDLM attention. For long-context inference, BDLM Mamba-H reaches 19.7x the throughput of full-sequence DiffuMamba-H at 65K tokens and 3.7x the throughput of BDLM attention at 262K, showing that Mamba hybrids are a potential long-context diffusion architecture.

Understanding and reasoning over long contexts has become a key requirement for deploying large language models (LLMs) in realistic applications. Although recent LLMs support increasingly long context windows, they often fail to use relevant evidence that is already present in the input, revealing a gap between context access and effective context utilization. In this work, we propose Recursive Evidence Replay as LLM Harness for Long-Context Reasoning (RECONTEXT), a training-free inference method for improving long-context reasoning. RECONTEXT uses model-internal relevance signals to construct a query-conditioned evidence pool and replays it before final generation while preserving the full original context. This recursive selection process separates evidence organization from answer generation without training, external memory, or context pruning. We also provide a theoretical analysis based on associative memory, which characterizes the context as a memory store, the question as a retrieval cue, attention as cue-trace association, and replay as trace reactivation. Experiments on eight long-context datasets with 128K context length show that RECONTEXT consistently improves evidence utilization across Qwen3-4B, Qwen3-8B, and Llama3-8B, achieving the best average rank on all three backbones. Code is available at https://github.com/Yanjun-Zhao/ReContext.

Linear-attention and state-space language models compress the prefix into a fixed-size recurrent state, yielding O(1) memory at the cost of a lossy exact memory: when many key--value associations compete, earlier facts are overwritten and needle recall degrades. Inspired by Complementary Learning Systems, we give linear attention a hippocampal complement. HOLA (Hippocampal Linear Attention) keeps the usual delta-rule state as a compressive memory and adds a bounded exact KV cache, forming a semiparametric test-time memory: the state models linearly compressible structure, while the cache stores associations that should not be forced through that state. The cache writes without a learned eviction module, keeping tokens with large beta * ||e||, the prediction residual actually committed to the state; a decoupled RMSNorm-gamma cache read then turns these exact KV pairs into sharp retrieval rather than soft averaging. At 340M parameters trained on 15B SlimPajama tokens, HOLA lowers Wikitext perplexity from 27.32 to 22.92 (-16.1%), below a full-attention Transformer++ (26.88), and improves LAMBADA perplexity from 30.95 to 30.26. It also achieves the best linear in-context retrieval and remains much more robust than GDN or a matched HOLA+recency cache on RULER needle-in-a-haystack recall out to 32k tokens (16x its training length).

Recurrent memory agents extend LLMs to arbitrarily long contexts by iteratively consolidating input into a fixed-size memory window. Despite their scalability, these agents exhibit a well-documented reliability problem: end-to-end performance degrades systematically as context length grows. We diagnose this failure by decomposing performance into two factors--memory capture and memory retention--and quantitatively confirm that retention is the dominant bottleneck. Retention collapses because existing designs maintain memory as a monolithic text block, forcing every update to risk overwriting previously retained content. Motivated by this diagnosis, we propose Multi-Head Recurrent Memory (MHM), a general, training-free framework that partitions memory into independent heads governed by a stage-wise select-then-update strategy. At each step, exactly one head is selected for update while the remaining heads are structurally shielded from overwriting, shifting the burden of retention from model behavior to architectural design. As a lightweight instantiation, we introduce Least-Recently-Updated MHM (MHM-LRU), which guarantees uniform head utilization with zero additional token overhead. Extensive experiments on long-context benchmarks show that MHM-LRU substantially improves both retention and end-to-end accuracy across the 100K--1M token range, where baselines degrade sharply. On RULER-HQA at 896K tokens, MHM-LRU improves the memory retention rate from less than 30% to 73.96%. These gains generalize across model families, scales, and task types, positioning architectural optimization as a practical and cost-efficient path toward reliable long-context recurrent memory.

In long-context use, large language models frequently synthesize answers from the meaning of a relevant context span rather than literally copy-pasting them. Identifying which attention heads perform this synthesis matters for interpreting long-context model behavior. Yet existing detectors miss these heads by construction: they reward heads whose attended token matches the generated token, a literal-copy criterion that captures where a head reads but not what it writes through its output-value (OV) circuit, the very mechanism that carries non-literal retrieval. We introduce Logit-Contribution Scoring (LOCOS), a write-aware detector that scores each head by the projection of its OV-circuit output onto the answer-token unembedding direction, contrasting needle and off-needle source positions in a single forward pass. Across three model families (Qwen3, Gemma-3, OLMo-3.1), mean-ablating the top LOCOS heads on the NoLiMa non-literal retrieval benchmark collapses ROUGE-L at lower head counts than prior attention-based detections; on Qwen3-8B, ablating 50 heads drives ROUGE-L from 0.401 to 0.000 while the strongest baseline still retains 0.292. The selected heads are retrieval-specific: parametric recall and arithmetic reasoning stay at baseline under the same ablation. On Qwen3-8B, the same ablation also drops MuSiQue from 0.55 to 0.08 and BABI-Long from 0.62 to 0.20, while a random-heads control stays within 0.05 of baseline.

Long-context LLM services now sustain prompts with hundreds of thousands to millions of tokens, making the key-value (KV) cache a first-order serving cost. Because the cache grows linearly with context length, it can exhaust GPU memory, force smaller batches, and reduce serving throughput. Prior KV cache compression techniques typically target only the sequence dimension or only the channel dimension, which leaves limited headroom as context windows scale. Compressing both dimensions promises higher memory reduction, but applying the two forms of compression directly leads to significant accuracy loss. This paper introduces MosaicKV, a dynamic two-D (dimensional) KV cache compression system for extremely long-context serving. MosaicKV uses dynamic two-D compression to address the accuracy challenge, exploiting the non-uniform importance distribution of elements within the KV cache. Instead of applying one compression pattern globally, MosaicKV identifies important elements for each KV vector and selects compression strategies at the granularity of KV cache segments. To address the performance challenge, where fine-grained sparsity and compression management overhead can offset the gains from compression, MosaicKV introduces compressed KV cache management. This mechanism uses underutilized GPU and CPU resources to maintain compressed KV caches and accelerate attention computation. Evaluation on an H800 GPU with multiple LLMs shows that MosaicKV delivers up to 16x attention speedup, 4.8x lower decode latency, and 7.3x higher throughput than the uncompressed baseline. At the same time, it reduces memory usage by 3x and incurs only 1.76% average accuracy loss on LongBench and RULER.

CLIP (Contrastive Language-Image Pre-training) has become a de facto paradigm for image-text alignment, but it struggles with long-context descriptions (>77 tokens) due to absolute positional encoding and pretraining on short captions. In long contexts, sentences are often reordered, summarized, or partially omitted. Although prior works extend CLIP with longer positional encodings, they often suffer from degraded image-text alignment under such text perturbations. We attribute this limitation to the Euclidean contrastive objective, which enforces strict one-to-one matching and lacks explicit mechanisms for modeling hierarchical relationships between global context and its constituent elements. To address this issue, we propose HyFL-CLIP, a hyperbolic fine-tuning framework that distills the well-established text-image alignment learned in Euclidean CLIP into hyperbolic space via cross-manifold similarity distillation, leveraging its geometry to capture hierarchical and entailment relations. Our method models hierarchical semantics by linking summarized token-wise features, long-context descriptions, constituent short textual components, and images, capturing part-whole relationships via hyperbolic entailment with Einstein midpoint aggregation. Experiments on diverse benchmarks, including long-context cross-modal retrieval, cross-modal retrieval with caption perturbations, intra-modality retrieval, and short-text cross-modal retrieval, show that HyFL-CLIP achieves more robust long-context understanding. In particular, it yields up to 19.5% improvement in long-text cross-modal retrieval under textual perturbations over the best prior method. We also show HyFL-CLIP can be seamlessly integrated into other model frameworks by applying it to Stable Diffusion XL (SDXL).

Large language models increasingly operate over long contexts, where the KV cache becomes a dominant memory bottleneck: its size grows linearly with sequence length and must be retained throughout decoding, making full GPU caching prohibitively expensive without compression. Existing KV cache compression methods struggle to balance efficiency with faithful context preservation. Token eviction discards information, while semantic grouping fixes compression decisions at prefill time; neither can recover token-level detail from a compressed span once it becomes relevant during generation. As a solution, we propose SeKV, a resolution-adaptive semantic KV cache that organizes context into entropy-guided semantic spans and stores them across a GPU-CPU memory hierarchy without discarding information. Each span keeps a lightweight summary vector on GPU for coarse routing and a low-rank SVD basis on CPU for on-demand token-level reconstruction. A trained zoom-in mechanism selectively expands query-relevant spans during decoding, enabling precise retrieval without materializing the full KV cache on GPU. SeKV enables adaptive token-level reconstruction while keeping the base LLM fully frozen and adding fewer than 0.05% trainable parameters. Across four benchmarks, SeKV improves over the strongest semantic compression baseline by 5.9% on average while reducing GPU memory by 53.3% versus full KV caching at 128K context. Code is available on https://github.com/AmirAbaskohi/SeKV.

Hybrid attention models improve long-context efficiency by retaining only a subset of full-attention layers and replacing the remaining layers with linear attention. However, the effectiveness of Transformer-to-hybrid conversion critically depends on which layers preserve full attention. Existing hybrid layer selection methods typically rely on heuristic strategies such as fixed placement patterns or layerwise scoring, implicitly treating layer importance as isolated and overlooking the interdependent layer effect under a global hybrid configuration. In this work, we formulate hybrid layer selection as a budget-constrained subset optimization problem. We further propose FlashMorph (Fast LAyer Selection for Hybrid MORPHing), an effective, efficient and scalable layer selection method for Transformer-to-hybrid conversion. FlashMorph first constructs a morphable model by equipping each full-attention layer with a converted linear-attention branch. It then freezes all model weights and jointly optimizes layerwise gates on synthetic long-context retrieval data, with a linearization regularization that encourages the model to rely on linear attention for efficiency. The learned gates are discretized under a preset full-attention budget to instantiate the hybrid architecture, followed by standard logits distillation and long-context finetuning. Extensive experiments show that FlashMorph discovers more effective hybrid configurations, preserves strong long-context recall and general benchmark performance while substantially reducing layer selection cost compared with existing layer selection methods, demonstrating its effectiveness, efficiency, and scalability.

Hierarchical Global Attention (HGA) is a drop-in replacement for dense causal attention in pretrained long-context transformers. HGA preserves the original checkpoint parameters: the pretrained $W_Q$, $W_K$, $W_V$, and $W_O$ projections remain unchanged, no calibration parameters are introduced, and no retraining is required. Applied to Qwen3-30B-A3B-Instruct-2507-FP8 on a single RTX~5090 (32GB), the patched model runs out of the box at a 64K-token context, where token-level K/V storage is not feasible on this hardware. Unlike previous sparse-attention methods, HGA performs hierarchical two-level routing. It first retrieves relevant chunks using compact RoPE-aware summaries and then refines the selection by routing only the most relevant groups before performing exact token-level attention. This hierarchical retrieval significantly reduces the number of fetched tokens while preserving exact attention over the retrieved token set, making RAM- and NVMe-backed storage practical. The full historical token K/V resides in host RAM or NVMe storage, while only a small routed working set is transferred to GPU memory during attention. Consequently, GPU memory consumption depends primarily on model weights and the routed working set rather than on the total context length. Across all tested context lengths (4K - 64K tokens), routed attention remains within approximately $0.01$--$0.02$ nats of dense attention while the sparsity used is just about 3%. These results suggest that the approximation introduced by hierarchical routing is small, and that the remaining quality gap is likely dominated by long-context positional encoding rather than by the routing algorithm itself.

In this paper, we aim to combine the advantages of existing sequence parallelism paradigms and overcomes their drawbacks, the most serious of which is the incapability to correctly compute causal attention on the hybrid-context packed sequences, in a stronger sequence parallelism framework. The practical technique of packing sequences for efficiently pretraining and fine-tuning large language models causes cross-contamination problem in attention computation, which can be effectively solved when no parallelism in the sequence length dimension is taken. However, in sequence parallelism, existing approaches either ignore the scenario of hybrid-context sequences or conversely sacrifice and limit parallelism degree for supporting the scenario. To this end, we innovatively propose an efficient Sequence-Aware Parallelism algorithm to conquer the obstacles of intensive tensor transmission and partial attention computation across multiple device groups. Our algorithm utilizes JIT (Just-In-Time) compilation to optimize the communication strategy of all device groups in NCCL level. Further, we integrate existing sequence parallelism paradigms into a Hierarchical Sequence-Aware Parallelism framework which benefits from our sequence-aware algorithm. We additionally elaborate on the memory and communication overhead management of the hierarchical framework to optimize its performance. Through multiple experiments, we demonstrate that our proposed approach outperform other state-of-the-arts sequence parallelism approches in multiple metrics.

Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV blocks relevant to each query, but it introduces a serialized selection-to-attention dependency that emerges as a new latency bottleneck. We present PRR, a speculate-reuse-repair runtime that exploits temporal locality in DSA selections to predict likely blocks, speculate the attention over them while selection is in flight, and incrementally repair missed blocks once the true selected set is known. PRR uses a lightweight EMA-based predictor, a profiling-guided speculation budget that keeps speculative work off the critical path, and a FlashAttention-based repair kernel that folds missed blocks into the partial attention state using online-softmax statistics. Across long-context benchmarks and representative DSA methods, PRR reduces per-token decoding latency by up to 40% while preserving downstream task accuracy. Github: https://github.com/Tianyu9748/Incremental_FlashAttention

The quadratic computational cost of traditional attention mechanisms poses a major bottleneck to the scalability and practical deployment of large language models (LLMs), particularly in long-context scenarios. To improve efficiency, existing approaches often enforce rigid structural constraints such as local attention windows. However, these strategies typically lead to substantial performance degradation on tasks requiring precise long-range recall. In this work, we propose MATCH, a scalable and efficient framework that augments sparsified attention mechanisms with dynamically integrated in-context information through an efficient retrieval system. Empirical results show that MATCH significantly improves the performance of sparse-attention models on both synthetic and real-world natural-language tasks. These findings highlight the versatility of MATCH as a general approach for enhancing in-context retrieval capabilities while maintaining the efficiency benefits of sparse attention architectures.

Extensive context has become the norm as Large Language Models (LLMs) are increasingly deployed in long-horizon tasks. The concern that increasing context length degrades model capabilities, known as context rot, has become a central issue for these applications. In this paper, we focus on deep search scenarios, aiming to investigate the rot phenomenon and its mitigation strategies. By evaluating four flagship open-source models across three benchmarks, we reveal a prevalent but unnoticed rot phenomenon: extensive context causes models to directly give up or prematurely provide uncertain answers, and this issue is exacerbated as the context grows. Through pruning experiments, we demonstrate the relationship between the accumulated context and the rot phenomenon. Furthermore, we investigate mitigating this issue through context management and post-hoc rejection sampling. For context management, we systematically evaluate seven different methods across three categories, based on performance, cost, and impact on context rot, providing clear guidance for strategy selection and usage. For rejection sampling, we develop a rot-aware filtering strategy and demonstrate its effectiveness across three aggregation methods. Finally, we show that these two approaches can be combined for further performance improvements.

Memory-Mediated Learning Architecture (MMLA) separates slow base parameters theta, a bounded numerical policy carrier Phi, and a bounded authoritative memory M. Predictive Dual-State Adaptation (PDSA) lets feedback update Phi while one problem remains active and lets a trusted lifecycle atomically commit one typed row or exact NULL. Later reasoning may read both states, but their writers, resets, rollback domains, and ledgers remain distinct. Realized futures supervise values only during training; deployment is causal and future-blind. We give conditional theory and falsifiable contracts for reasoning-time updates, completed-segment consolidation, predictive admission, authoritative memory, and dual-state attribution. Assumptions, counterexamples, capacity and cost ledgers, recovery duties, and identifying experiments are explicit; these are not implementation guarantees. Controlled studies show exact lifecycle execution on 300/300 held-out records for each of three seeds, calibrated retrieval gains over frozen-hidden dense and BM25 baselines, and exact typed anchor-filler transport on 240/240 held-out records per seed. These validate restricted components, not natural-language memory management or complete PDSA. Post-training studies retain positive and negative evidence. Later protocols obtain restricted readout progress, but a nine-trajectory comparison finds that matched latent readout, a full-width bridge, and bridge plus frozen text-teacher alignment all fail continuous-event qualification across both task families. A separately adapted text reference and restoration checks pass. At the September 13, 2026 evidence cutoff, no strict policy-only reasoning-time-training effect, predictive-admission oracle margin, learned future-blind admission policy, or policy-by-memory factorial advantage is established.

Sparse autoencoders (SAEs) have become a useful tool for extracting interpretable features in language models. However, standard SAE architectures operate on individual token activations, meaning that the number of active features scales linearly with context length, and studying long model transcripts becomes difficult. We introduce turn-averaged SAEs, which represent a single Human or Assistant turn with a fixed number of features by learning to reconstruct the average model activation across the turn. We find that turn-averaged features describe a single turn's high-level characteristics more completely than per-token features when judged by an LLM. We also demonstrate that turn-averaged SAEs greatly simplify common downstream uses of SAEs like attribution graphs. Broadly, turn-averaged SAEs make interpretability techniques practical at long context lengths.

Hybrid attention models that mix full and sliding-window attention across layers offer a promising approach to efficient long-context inference, but the critical question of \emph{which layers} should retain full attention remains unsolved. Existing methods use either fixed periodic patterns or attention-based heuristics that may not capture what matters for downstream accuracy. We propose NLL-guided layer selection, a training-free method that directly measures each layer's importance by computing the negative log-likelihood degradation on answer tokens when that layer uses sliding-window instead of full attention. On LongMemEval with Qwen3-4B, our method achieves 64.6% accuracy using only 1/4 full-attention layers, matching the 1/2-FA periodic baseline (65.0%) while halving the computational budget. NLL-guided selection outperforms the SWAA-reported periodic 1/4-FA baseline by 10.4 percentage points and a matched LightTransfer-style baseline by 26.4 percentage points. De-confounding analysis shows the signal is consistent with long-range attention needs rather than generic layer sensitivity. The method requires only $\sim$15 minutes of one-time calibration, advancing the efficiency-accuracy Pareto frontier for long-context LLM deployment.

Reasoning capability has advanced rapidly in large language models (LLMs), leading to an increasing size of key-value (KV) cache in both prefilling and decoding stages. Existing KV cache compression methods mainly rely on attention weights to estimate token importance. While attention effectively captures contextual relevance, it overlooks complementary information-theoretic signals related to predictive uncertainty and token informativeness. In this paper, we revisit token importance from a forward-looking perspective and introduce \textit{Forward Influence}, a metric that measures how compressed tokens affect future contexts. Our analysis reveals that tokens selected by attention scores mainly influence nearby contexts, whereas tokens associated with high predictive uncertainty exhibit substantially stronger influence on distant future contexts. Based on the observation, we propose \textbf{InfoKV}, an entropy-aware KV cache compression framework that incorporates information-theoretic signals. It combines token-level predictive uncertainty with layer-wise representation evolution and integrates the resulting entropy scores with attention scores during reasoning. Experiments on long-context reasoning benchmarks with Llama-3.1, Llama-3.2, and DeepSeek-R1 demonstrate that InfoKV consistently outperforms existing attention-based KV compression methods in both long prefilling and decoding scenarios.

Delta-rule linear attention improves recurrent memory updates by correcting what is already stored at the current write address before writing new content. However, the active correction is still anchored to that same write address. As a result, stale information stored at a different address cannot be actively removed before new content is written elsewhere. We propose Erase-then-Delta Attention (EDA), a memory update rule that decouples where to erase from where to write. The key insight is that recurrent memory models should not only correct the current write, but also selectively suppress outdated memory at an independently chosen address. Concretely, our method first applies a targeted erase step along a learned erase direction, and then performs the standard delta-style corrective write along the current write direction. This preserves the corrective behavior of delta-rule updates while expanding their memory-management capacity. Language-model pretraining experiments across dense 2.5B and MoE 25B-A2.8B model families show that EDA performs best in both settings. The gain persists after 80B-token long-context midtraining of the MoE models, where EDA also performs best in long-context evaluations from 4k to 128k contexts. A compact update analysis and memory-state probes suggest why: EDA keeps the delta-rule corrective write intact while allocating an additional cleanup path most strongly when passive decay is weak. These results suggest that recurrent memory models should decide not only what to write, but also what stale information to erase and where.

While parameter-efficient fine-tuning (PEFT) typically targets attention projectors, its efficacy for tasks requiring sequential state accumulation remains under-explored. We examine if PEFT for such tasks can benefit from state space model (SSMs) adapters, and if MLP blocks are better injection sites. We introduce Hankel Reduced order Model (HRM) adapter, an SSM-based residual module initialized via Balanced Truncation of empirical Hankel Grammians. By leveraging the time-invariance of the system matrix $\bar{A}$, HRM enables an exact FFT-based parallel scan, achieving computational parity with LoRA across all context lengths. In iso-parametric evaluations on Mistral-7B (8.4M trainable parameters), HRM outperforms LoRA variants on LongBench tasks, including QuALITY (+34.8% relative accuracy) and QMSum (+71.6% relative ROUGE-1). HRM further demonstrates consistent superiority across 18 configurations of synthetic state-tracking (DFA, Parity) and character-level language modeling (enwik8). Gate analysis reveals that HRM adapters effectively learn to modulate recurrence, providing a robust architectural alternative to low-rank adaptation for long-context sequence modeling.

Modern large language models based on softmax scaled-dot-product attention are constrained by their training sequence length: as the key-value sequence grows, softmax probability mass can dilute across a wider distribution, inducing activation shift and long-context performance collapse. Moreover, long-context language modeling faces a structural tension: a sliding-window attention core maintains a bounded local representation and low perplexity but is blind to long-range dependencies, while full-context attention preserves global recall but suffers from out-of-distribution perplexity explosion. To resolve these limitations, we introduce ATMA, a hybrid convolutional-attention architecture that integrates a novel three-channel attention mechanism. ATMA factorizes the attention mixing step into: (1) a count-blind, unit-vector direction channel, (2) a bounded magnitude channel driven by the participation ratio of effective matches over an extreme-value-corrected null sink, and (3) a long-term recurrent compression memory optimized via a gated-delta fast-weights rule. Neither the Polar Attention core nor the recurrent memory is sufficient alone; their combination enables monotonic perplexity reduction and high-fidelity long-range retrieval simultaneously. We evaluate ATMA using a 120-run factorial ablation sweep, demonstrating that the combined Polar + memory model maintains induction needle-in-a-haystack retrieval accuracy above 90% out to 64K tokens (32 times the training length of 2K) while its document perplexity improves monotonically, outperforming softmax-based memory baselines which collapse at extreme context lengths. Code: https://github.com/kreasof-ai/atma

PaTH Attention showed that replacing RoPE's position-indexed rotations with accumulated data-dependent Householder reflections yields strong length extrapolation, though performance degrades at extreme context lengths. We ask whether this depends on Householder-specific structure or reflects a general property of accumulated transformations along source-to-query paths. We study a simpler variant keeping RoPE's block-diagonal SO(2) rotations but replacing position-indexed angles with accumulated token-dependent ones. It shows the same pattern: improved extrapolation then degradation at long contexts. We prove the result extends to accumulated orthogonal transformations satisfying certain regularity conditions: their products become incoherent after finitely many steps, suppressing attention to distant tokens. Accumulated rotations of queries and keys create a finite mixing window independent of context length; per-token suppression learned in training transfers unchanged to any evaluation length, and high-dimensional concentration produces a score gap suppressing far tokens while near-route transport preserves the target signal. Conversely, a lower bound shows accumulated rotations must eventually degrade: as the far set grows, no rotations preserve the near signal without explicit far-mass control. For SO(2) rotations, rotating values too makes residual far contributions combine incoherently, extending the range. Controlled experiments support these predictions: random accumulated rotations substantially improve extrapolation over RoPE, learned token-dependent rotations maintain near-training-length perplexity far beyond the training context, and rotating values helps over queries and keys alone. Rotation-only models still degrade at extreme lengths, while ALiBi stays length-stable, consistent with the need for far-mass control.

While speculative decoding improves inference throughput for multi-batch long-context Large Language Models (LLMs), its efficiency is often limited by a verification bottleneck where Key-Value (KV) cache loading dominates latency. Existing compression methods fail in this regime: static eviction incurs accuracy loss due to saliency shift, while dynamic selection introduces prohibitive computational overhead during the verification path. We propose Dustin, a sparse verification framework designed for long-context speculative decoding. Dustin integrates lookahead signals from the draft model with historical attention from the target model to identify critical tokens with high fidelity across multi-step verification windows. To reduce recomputation latency, this approach further employs a sparse estimation scheme that restricts importance scoring to a minimal subset of attention heads. Evaluations on PG-19 and LongBench with Qwen2.5-72B demonstrate that Dustin achieves a 27.85x speedup in self-attention and a 9.17x end-to-end decoding speedup at a 32k sequence length, all with negligible accuracy degradation.

Contextual entrainment, which is a newly discovered phenomenon in large language models (LLMs), refers to the tendency of a model to assign higher probabilities to tokens that appear in its context. In this work, we extend this phenomenon from the token level to the sentence level by examining the per-token mean log-probability of a sentence instead of the probabilities of individual tokens. We investigate sentence-level contextual entrainment across 26 LLMs from seven families and two datasets, which cover both subjective and objective tasks. We find that sentence-level contextual entrainment exists. This means that the sentences in the prompt (even if they are counterfactual statements) can significantly increase their probability during model inference time. As the model size increases, contextual entrainment gradually decreases. We also find that contextual entrainment is controlled by 2% to 4% of the attention heads. Turning off these attention heads can effectively mitigate contextual entrainment without hurting the model's performance.

Large language models (LLMs) are typically pretrained on short sequences and then extended to work on longer sequences with additional training. However, such LLMs still struggle to further generalize to very long sequences. We propose Randomized YaRN, a training method that improves length generalization by combining YaRN-based positional extrapolation with randomized positional encoding and a length curriculum. During training on short context data, tokens are assigned YaRN positional encodings sampled from a larger position range, exposing the model to out-of-distribution positional representations even on short-context inputs. We evaluate Randomized YaRN on three challenging long-context reasoning benchmarks, BABILong, Multi-Round Coreference Resolution (MRCR), and LongBench v2. When training on data with short context, Randomized YaRN consistently improves reasoning performance on context lengths from 16K to 128K and outperforms standard fine-tuning, with the largest gains appearing at far out-of-distribution lengths. Our results suggest that progressively exposing models to OOD positional distributions provides an effective recipe for generalizable long-context reasoning.

Backpropagation-trained dense neural networks are powerful function approximators, but they couple learning across many parameters and can overwrite previous associations when tasks conflict. This paper describes Local Pheromone Network, a small research prototype for sparse, local, manually updated neural networks. In Local Pheromone Network, each output unit reads only a fixed local neighborhood of input units subject to geometric distance and molecular-tag compatibility. Each synapse stores a weight, a short-term pheromone trace, a long-term pheromone trace, and an optional consolidation state. Training does not call automatic differentiation. Instead, every layer performs a pheromone-weighted Hebbian-style update on a budgeted subset of local synapses selected from local error and co-activity. The update budget adapts online: it shrinks when loss improves and expands toward recently active neighborhoods when loss worsens. Optional mechanisms add structural plasticity, local replay, output masks for partitioned learning, and a target-free local contrastive step. We present the implementation, learning rule, and preliminary experiments on synthetic regression, partitioned memory, conflicting memory, consolidated conflict, structural plasticity, replay, and a synthetic long-context hybrid memory task. The prototype learns local linear rules, preserves partitioned memories through tags and masks, reduces forgetting under consolidation, and uses replay under conflict.

Time series forecasting is a fundamental machine learning task. Recent work has explored Large Language Models (LLMs) for this purpose due to their strong generalization, pattern recognition, and zero-shot or few-shot capabilities. Despite their suitability for long-context learning, LLMs face challenges in multimodal settings: they lack calibrated probabilistic modeling for non-text data and struggle to align heterogeneous representations. To address these issues, we propose a new framework Diffusion-LLM that integrates a conditional diffusion model into an LLM-based forecasting pipeline. This joint design enables learning the conditional distribution of future data while improving semantic alignment in a shared latent space. We evaluate Diffusion-LLM on six long-term forecasting benchmarks, including ETT, Weather, and ECL. Our method consistently outperforms existing LLM-based baseline, achieving notable gains in ultra-long-term and few-shot forecasting and demonstrating the value of distribution-aware regularization for enhancing robustness and generalization in time series LLMs.

Long contexts have become standard in pretrained LLMs, yet they remain expensive to run: prefill compute grows quadratically with sequence length, and every decode step re-reads a key-value cache that grows linearly with it. Sparse attention cuts these costs by attending only to a relevant subset of past tokens, but selecting that subset is itself expensive. We present SpotAttention, a lightweight selector that attaches to a frozen pretrained transformer and learns by KL distillation to estimate its attention distribution. The selector picks the top-K keys each query attends to, and because its estimate is a calibrated distribution, a dual top-p rule reads the per-query, per-layer budget directly from it. Across Qwen3 (dense, 4B-32B) and Qwen3.5 (hybrid linear/full attention, 4B-9B), SpotAttention matches dense accuracy at contexts up to 128K tokens, eight times the training length. Decode at L=128K runs 3.9x faster than FlashAttention and 1.8x faster than Twilight, the strongest training-free baseline. Quantizing the selector's K-cache to INT4 or FP4 microscale shrinks it 3.5x at no accuracy cost.

Existing embedding models are inherently static: they encode text segments in isolation, ignoring their surrounding context and temporal order. This paper introduces EvoEmbedding, a novel embedding model that generates evolvable representations for retrieval. It is tailored for long-context scenarios, where information is dynamic, sequential, and requires continuous state tracking. Our design is simple: EvoEmbedding maintains a continuously updated latent memory as it sequentially processes inputs, and uses it alongside the raw content to jointly generate evolvable embeddings. Consequently, for the same query, our model adapts its representation to retrieve distinct targets based on the evolving context, going beyond static semantic search. To equip the model with this capability, we construct EvoTrain-180K, a diverse dataset for the joint optimization of latent memory and retrieval. Furthermore, we introduce a memory queue to prevent representation collapse during recurrent encoding, alongside segment-batching techniques that tackle significant length variance and accelerate training by 3.8$\times$. Extensive experiments show that our model not only outperforms larger-scale specialists (e.g., Qwen3-Embedding-8B and KaLM-Embedding-Gemma3-12B) across a range of long-context retrieval benchmarks, but also generalizes well to downstream tasks (e.g., personalization) with contexts 10$\times$ longer than its training window. Notably, EvoEmbedding seamlessly integrates into agentic workflows to boost performance. For instance, a naive RAG pipeline equipped with our model surpasses dedicated agentic memory systems. Project Page: https://clare-nie.github.io/EvoEmbedding/.

Long-form narratives are challenging for long-context models because their structure is implicit: events, characters, and plotlines interact across hundreds of pages without the explicit cues that guide navigation in structured documents. We address this by constructing storyline trees, hierarchical representations that organize narratives from global themes and major plotlines to fine-grained events. We first segment chapters into contiguous narrative segments, or scenes, and use them as the basic units for tree construction. We then infer storyline trees through complementary top-down and bottom-up procedures that derive, refine, cluster, and summarize storylines at multiple levels of abstraction. We showcase the utility of this representation for question answering: storyline trees enable adaptive retrieval, allowing models to iteratively inspect high-level narrative structure and retrieve scene-level evidence on demand. Experiments on three long-context narrative QA benchmarks show that adaptive retrieval outperforms strong baselines, including post-trained long-context models and agentic chunk-based methods. Ablations confirm that scenes are more effective basic units than chapters or generic segmentation, and that gains persist under matched retrieval budgets

Document-grounded assistants built on large language models are increasingly used in high-stakes, knowledge-intensive work. Their usefulness, however, may depend on how evidence is allocated before generation. We investigate such a claim by comparing two grounding architectures: (a) retrieval-augmented generation (RAG) that retrieves a few relevant passages, and (b) long-context prompting, which loads the whole document collection in context. We view these as two regimes of "epistemic access" on an accuracy--cost frontier. We use "epistemic accuracy" to capture model correctness that depends on having the right evidence. We posit that broader access (via long context) can increase it, but with a "token tax" (i.e., a substantial increase in cost due to larger input token consumption). We probe this framing with a case study in manufacturing safety training. Using an expert-validated benchmark, we evaluate 972 answers across three machines, two small language models, and three retrieval/in-context prompting approaches. Long-context prompting achieved the highest correctness (73.1% vs. 65.4% for semantic RAG), but at 26 times the per-query token cost. We interpret this gap as the token tax of broader evidentiary access. We carefully discuss the implications of our findings for resource-constrained organizations.

The quadratic complexity of attention poses a critical bottleneck for long-context processing, spurring interest in hybrid attention designs. Most open-source hybrid models adopt a layer-wise strategy. Yet, prior work has noted the inherent difficulty of integrating Linear Attention (LA) with Full Attention (FA), suggesting that the design space of attention hybridization remains underexplored. To probe this space, we conduct interpretability analysis and observe that layers exhibit block-wise functional similarity, while individual heads within the same layer display distinct functional specialization despite sharing input features. This head-level heterogeneity suggests that the head dimension provides a natural and principled granularity for fusing heterogeneous attention signals. Building on this insight, we introduce HydraHead, a novel architecture that hybridizes FA and LA along the head axis. HydraHead features two key innovations: (1) an interpretability-driven selection strategy that identifies retrieval-critical heads and preserves FA only for them, and (2) a scale-normalized fusion module that reconciles the distributional gap between FA and LA head outputs. By leveraging a three-stage transfer pipeline with parameter reuse and distillation, we achieve high-performance hybrid models with minimal training overhead. Under a unified training setup, HydraHead outperforms other hybrid designs in long-context tasks while maintaining strong general reasoning. With interpretability-driven head selection, it matches a 3:1 layer-wise hybrid's long-context performance at a 7:1 LA-to-FA ratio. Crucially, trained on only 15B tokens, HydraHead achieves over 69% improvement over the baseline at 512K context length, approaching Qwen3.5, a leading model of comparable size with a native context length of 256K. This highlights the significant scaling potential of head-level hybridization.

Post-hoc context erasing over the KV cache is challenging because a local edit has a global consequence: once a span has been processed, its influence propagates into the cached states of all subsequent tokens. This issue arises naturally in long-context LLM applications, where stale retrieved facts, incorrect tool observations, retracted user preferences, or harmful prompt injections may be identified only after prefill. Exact erasing must then recompute all tokens after the deleted span, making its computational cost depend on suffix length rather than erased-span length. We introduce KVEraser, a learned KV-cache editing method for efficient localized context erasing. Given a processed context and a span to remove, KVEraser replaces only the KV states of the erased interval with learned steering states while reusing the remaining cache unchanged. To learn a transferable erasing mechanism, we build a two-stage training pipeline: generic span-neighbor pre-training teaches the eraser to suppress the influence of the erased span, while task-specific fine-tuning adapts this capability to downstream scenarios. Experiments show that KVEraser nearly matches full recomputation in post-erasure performance on in-domain tasks across 1K--32K context lengths, while its latency increases by only 24% compared with a 17.6x increase for full recomputation. KVEraser also generalizes to unseen long-document QA tasks with harmful factual distractors, achieving the best performance among approximate baselines with a 3--4x speedup over full recomputation.

Pretraining language models with extended context windows enhances their ability to leverage rich information during generation. Existing methods split input sequences into chunks, broadcast them across multiple devices, and compute attention block by block which incurs significant communication overhead. While feasible in high-speed clusters, these methods are impractical for decentralized training over low-bandwidth connections. We propose a compression method for communication-efficient context parallelism in decentralized settings, achieving a remarkable compression rate of over 95% with negligible overhead and no loss in convergence. Our key insight is to exploit the intrinsic low-rank structure of activation outputs by dynamically constraining them to learned mixtures of subspaces via efficient reparameterizations. We demonstrate scaling billion-parameter decentralized models to context lengths exceeding 100K tokens on networks as slow as 300Mbps, matching the wall-clock convergence speed of centralized models on 100Gbps interconnects.

Modeling long-range dependencies remains a central challenge in natural language processing. Transformer architectures achieve strong performance via self-attention but scale quadratically ($O(N^2)$) with sequence length, while State Space Models (SSMs) scale linearly ($O(N)$) but suffer from a selective recall bottleneck, struggling to retrieve precise information from compressed states. This creates a fundamental tradeoff between efficiency and perplexity. To tackle these challenges, we propose the \textit{Parallel Hybrid Architecture (PHA)}, which runs Gated State Spaces (GSS), Grouped Query Attention (GQA), and Feed-Forward Networks (FFNs) as independent parallel branches fused by a learnable mixing mechanism. Instead of forcing SSMs to approximate attention or serializing the two paradigms, PHA allows each branch to specialize: GSS captures global context, while attention performs selective retrieval, with FFN providing complementary processing. On WikiText-103, PHA achieves 16.51 PPL at 125M parameters, outperforming Hedgehog (16.70) and H3-125M (23.70). Scaling to 180M parameters yields 16.42 PPL, which gives comparable results with the pure attention baseline while delivering 24% higher throughput and up to 40% lower memory usage at long contexts. On OpenWebText, our 125M model achieves 19.72 PPL, outperforming standard Transformers (20.60) and GSS hybrid baselines (19.80). These results demonstrate that separating sequence modeling paradigms into parallel specialists enables Transformer-level perplexity with substantially improved efficiency for long-context language modeling.

Modern language models increasingly adopt hybrid architectures that combine full attention with efficient attention modules, such as sliding-window attention (SWA) and recurrent sequence mixers. However, how these efficient modules shape model capabilities remains poorly understood. To address this gap, we conduct a systematic analysis across hybrid architectures from three perspectives: scaling behavior, mechanism analysis, and architecture design. First, from a scaling perspective, we find that efficient-attention design primarily affects how fast long-context capability emerges, while different hybrids eventually converge to comparable long-context performance under sufficient training. Second, mechanistically, we show that long-range retrieval is mainly carried by full attention, whereas efficient attention shapes its optimization trajectory. This explains a counter-intuitive phenomenon we call Large-Window Laziness: larger SWA windows can delay the formation of retrieval heads in full-attention layers. Third, guided by this mechanism, we show that applying NoPE to only the full-attention layers of a small-window SWA hybrid substantially improves long-context performance with negligible impact on short-context performance.

Long-context language modeling requires not only extending context windows but maintaining coherent understanding of entity states and relationships across thousands of tokens -- a challenge that semantic similarity alone cannot address. KGERMAR addresses this by constructing dynamic, context-specific knowledge graphs from input text during inference, enabling domain-adaptive retrieval that leverages both semantic similarity and explicit entity relationships. The framework performs real-time entity and relation extraction to build contextual knowledge graphs, then integrates graph-structural embeddings with textual semantics through a multi-component memory architecture. Three memory banks -- contextual, semantic, and structural -- are maintained with retrieval signals fused via learned weights to capture both surface-level semantics and deeper relational patterns. Evaluated on SlimPajama (84.7K training examples), WikiText-103 (4,358 examples), PG-19 (100 examples), and Proof-pile (46.3K examples), KGERMAR achieves up to 8.5% lower perplexity and 2--2.5x better memory efficiency than memory-augmented baselines across context lengths from 1K to 32K tokens, with superior in-context learning performance across five NLU tasks. The dynamic knowledge graph construction approach advances memory-augmented language modeling by enabling domain-specific knowledge representation that adapts to input contexts rather than relying on fixed knowledge bases.

While Large Language Models (LLMs) have advanced open-domain dialogue systems, maintaining long-term consistency remains a challenge due to inherent limitations in long-context reasoning and the inefficiency of processing extensive raw text. Existing approaches typically rely on either unstructured memory storage, which is prone to information loss, or computationally expensive LLMs that incur high latency. To address these limitations, we propose G-Long, a graph-enhanced framework that utilizes a fine-tuned small Language Model (sLM) for structured triplet extraction and associative retrieval, significantly reducing operational costs. Furthermore, we introduce the novel attention-aware importance scoring mechanism that leverages the intrinsic cross-attention signals of a T5 summarizer to identify salient memories. Extensive experiments across diverse benchmarks demonstrate that G-Long achieves state-of-the-art performance in both response generation and memory retrieval, yielding performance gains of up to 9.8% in response quality on MSC and 40.8% in retrieval recall on LME, while significantly minimizing computational overhead.

Modern conversational agents condition on an ever-growing dialogue history at each turn, incurring redundant attention and encoding costs that grow with conversation length. Naive truncation or summarization degrades fidelity, while existing context compressors lack cross-turn memory sharing or revision, causing information loss and compounding errors in long dialogues. We revisit the context compression under conversational dynamics and empirically present its fragility. To improve both efficiency and robustness, we introduce Context-Driven Incremental Compression (C-DIC), which treats a conversation as interleaved contextual threads and stores revisable per-thread compression states in a single, compact dialogue memory. At each turn, a lightweight retrieve, revise, and write-back loop shares information across turns and updates stale memories, stabilizing long-horizon behavior. In addition, we adapt truncated backpropagation-through-time (TBPTT) to our multi-turn setting, learning cross-turn dependencies without full-history backpropagation. Extensive experiments on long-form dialogue benchmarks demonstrate superior performance and efficiency of C-DIC; notably, C-DIC shows stable inference latency and perplexity over hundreds of dialogue turns, supporting a scalable path to high-quality dialogue modeling.

Chain-of-thought (CoT) supervised fine-tuning (SFT) is widely adopted to improve reasoning ability, yet we find that it systematically degrades long-context recall in hybrid linear-attention models. Across architectures including HypeNet and Jet-Nemotron, retrieval performance on Needle-In-A-Haystack (NIAH) deteriorates substantially after CoT-SFT, and the degradation becomes more severe under harder retrieval settings and longer context windows. For example, HypeNet-9B on NIAH-S2@256K decreases from $67.2\%$ to $9.4\%$. We attribute this to CoT-SFT biasing attention gradients toward short-range patterns, disrupting query-key projections ($W_Q, W_K$) that are responsible for long-range routing. Motivated by this observation, we propose QK-Restore, a training-free method that restores only $W_Q$ and $W_K$ from the pre-SFT checkpoint while preserving all other post-SFT parameters. We further introduce a Procrustes variant to balance routing preservation and reasoning adaptation. Across architectures, QK-Restore consistently restores long-context capability at zero training cost while preserving reasoning performance; for instance, on HypeNet-5B it improves S3@256K from $65.4\%$ to $76.4\%$ while maintaining strong reasoning performance.

Pre-trained language models (PLMs) have achieved strong performance in keyphrase extraction (KPE), largely due to their ability to generate rich contextualized representations. However, long-document KPE remains challenging because salient keyphrase evidence may be scattered across distant document sections that cannot be jointly captured within the limited context window of most PLMs. Although long-context large language models (LLMs) can process broader textual contexts, their computational cost limits their practicality for efficient and high-throughput KPE. To overcome this limitation, we propose an attention expansion mechanism that augments PLM token representations with information from surrounding out-of-context chunks using pre-trained word embeddings. The proposed mechanism expands the effective contextual scope of PLM-based KPE models without requiring full-document attention or expensive LLM-based inference. We evaluate our approach across five PLM backbones, including general-purpose, scientific, task-specific, and long-context encoders, using two training regimes and five benchmark corpora from scientific and news domains. Experimental results demonstrate that attention expansion consistently enhances KPE performance across all evaluation settings, outperforming state-of-the-art models and yielding notable improvements in F1 score. The improvements extend to domain-specific, task-specialized, and native long-context models, showing that the proposed mechanism provides complementary information rather than merely compensating for limited input length. These results establish attention expansion as an efficient and effective strategy for long-document KPE.

The scalability of Large Language Models (LLMs) to long contexts is fundamentally constrained by the quadratic complexity of standard attention, motivating the adoption of linear attention mechanisms with sub-quadratic cost. To improve representation capacity under long contexts, recent approaches organize memory in a multi-state manner. However, existing multi-state linear attention methods rely on fixed state merging policies that cannot adapt to dynamically varying token importance, irreversibly obscuring critical tokens and causing severe error accumulation over long sequences. To address this limitation, we propose DLA, a dynamic memory modeling framework for multi-state linear attention. DLA introduces (i) Information-Aware Dynamic State Merging, which adaptively determines state boundaries based on token-level information variation, preserving high-resolution representations around semantic transitions while aggressively summarizing stable regions, and (ii) Capacity-Bounded Memory Modeling, which maintains a fixed-size, chronologically ordered state cache by selectively merging adjacent low-information states to control memory growth with minimal information loss. We pre-train DLA on two different linear attention models and evaluate on 16 datasets across three categories. Experimental results demonstrate the superiority of DLA over state-of-the-art.

Rotary Position Embeddings (RoPE) make attention scores position-relative but leave the value pathway position-blind: the message sent by a value token is the same regardless of its distance from the query. We propose RoVE, a parameter-free modification that makes values position-sensitive by rotating them simultaneously with keys, and show that it turns RoPE attention into attentive convolution. This new perspective unifies several independent formulations of the same operation across computer vision, robotics, and modern LLM architectures. Trained 124M and 354M GPT-2 models show consistent empirical gains over RoPE on few-shot in-context learning, out-of-distribution perplexity, and long-context retrieval, with the clearest improvements on tasks that require long-range aggregation.

Transformers achieve strong language modeling performance by providing direct token-to-token communication paths, but causal self-attention scales quadratically with context length. Recurrent and state-space models reduce this cost, yet compress history into sequentially updated fixed-size states. This paper studies a third primitive: a parallel content-addressed memory over causal successor records. The proposed Parallel Causal Associative Field (PCAF) writes local records from a context window into hash buckets, retrieves a bounded candidate set for the current query, forms a sparse cache distribution over successor tokens, and mixes that cache with a parametric local language model through a learned gate. The resulting model maintains sparse long-context access while avoiding a single fixed recurrent state bottleneck. We evaluate PCAF under full autoregressive pretraining on WikiText-103 and PG-19 using a distributed Google Cloud TPU v4-32 pod. At 303M parameters and context length T = 2048, PCAF-semantic reaches 36.31 perplexity on WikiText-103 and 52.45 perplexity on PG-19, compared with 47.49 and 53.84 for a matched dense Transformer. PCAF-semantic simultaneously processes 0.61-0.62M tokens/s across the TPU pod, versus 0.43M tokens/s for dense and local attention baselines. Supporting 41M-parameter multi-seed sweeps and single-GPU component ablations show that the associative cache, retrieval capacity, and learned gate materially affect the speed-quality trade-off.

Long-context language model inference is bottlenecked by memory, as the KV cache grows with context length. Recent techniques to compress the KV cache fall short: they either degrade model quality substantially or require considerable time and compute to compress a single long prompt. Furthermore, many methods require the input to fit within the target model's context window, and are generally incompatible with modern production inference engines. Encoder-decoder compressors, which map a long token sequence to a shorter sequence of latent embeddings consumed by a decoder, are an appealing alternative in principle. However, existing approaches are not competitive with KV cache compression on the accuracy-efficiency frontier. In this work, we revisit encoder-decoder compression and close this gap. We first perform an architecture search, pre-training many variants from scratch to determine how best to design and train encoder-decoder compressors. Guided by our findings, we continually pre-train a family of 0.6B-encoder, 4B-decoder models on over 350B tokens each, at compression ratios of 1:4, 1:8, and 1:16. We introduce Latent Context Language Models (LCLMs), a family of compressors that improve the Pareto frontier across general-task performance, compression speed, and peak memory usage. We demonstrate that LCLMs serve as efficient backbones for long-horizon agents, letting the agent skim through a compressed long context and adaptively expand relevant segments on demand.