arXiv:2607. 07740v1 Announce Type: cross Abstract: 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.
arXiv:2607. 19358v1 Announce Type: new Abstract: Recent advances in long chain-of-thought reasoning models such as DeepSeek-R1 have led to increasingly longer inference context lengths under the test-time scaling paradigm.
arXiv:2608. 02947v1 Announce Type: new Abstract: The attention score with rotary position embeddings (RoPE) decomposes exactly into a sum over its 2D-rotation frequency pairs, and each pair's wavelength limits how far it can discriminate position.
arXiv:2512.16391v2 Announce Type: replace-cross Abstract: Attention is the dominant source of latency during long-context LLM inference, an increasingly popular workload with reasoning models and RAG...
The paper introduces NAMOH, a native sparse attention mechanism that activates only a subset of heads per token, allowing each head to attend to a limited subsequence of tokens. By scaling the number of heads while keeping the active heads per token fixed, the method shortens head histories and reduces key‑value access without increasing overall storage. Experiments demonstrate that NAMOH can outperform fully activated models with the same parameter count and enable more efficient long‑context inference than smaller dense models.
The paper reports that chain‑of‑thought (CoT) supervised fine‑tuning (SFT) improves reasoning but systematically harms long‑context recall in hybrid linear‑attention models such as HypeNet and Jet‑Nemotron. Retrieval performance on the Needle‑In‑A‑Haystack benchmark drops dramatically after CoT‑SFT, especially with harder settings and longer contexts. The authors introduce QK‑Restore, a training‑free method that reinstates the query‑key projection matrices from the pre‑SFT checkpoint, which recovers long‑range recall while preserving reasoning gains.
The attention score with rotary position embeddings (RoPE) decomposes exactly into a sum over its 2D-rotation frequency pairs, and each pair's wavelength limits how far it can discriminate position. Aligned with this structure, we propose the per-RoPE-wavelength distance window: it prunes the query--key inner-product terms beyond a wavelength-proportional distance.
BF1 is a deterministic block‑aligned dyadic sparse‑attention retrofit designed to reduce the cost of causal attention in long‑context transformers. It combines a small exact local neighborhood, a global first block, and logarithmically spaced historical blocks, achieving O(n log n) token interactions per layer with O(log n) communication depth. On an NVIDIA RTX PRO 6000 Blackwell GPU, BF1 outperforms dense attention for 2K–4K tokens and delivers up to a 10.91× prefill speedup at 32K tokens, while retrofitting eight of 28 Qwen3‑0.6B layers reduces first‑token latency by up to 15.3% at 32K tokens and yields the lowest perplexity among compared sparse and dense training protocols.
arXiv:2602.08005v2 Announce Type: replace-cross Abstract: Efficient long-context inference faces two coupled bottlenecks: KV-cache memory grows linearly with context length, while attention computati...
arXiv:2609.36835v1 Announce Type: new Abstract: Long-context large language model inference is bottlenecked by KV caches that grow linearly with sequence length. This burden is especially severe for...
arXiv:2607. 01523v1 Announce Type: cross Abstract: Recurrent memory agents extend LLMs to arbitrarily long contexts by iteratively consolidating input into a fixed-size memory window.
CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling) is a new method for long-context LLM inference that replaces costly quadratic attention prefilling with a dynamic, input-adaptive sparse routing scheme. It introduces a structural proxy, C_struct, to directly read routing decisions from the proxy attention map, eliminating the need for pooled matrix multiplication and KL divergence. Additionally, CRISP addresses the post-softmax mass cliff by using a sink-aware threshold based on the noise floor, theoretically reducing background noise accumulation to O(n). Empirical results on InfiniteBench, RULER, and LongBench show that CRISP outperforms existing sparse methods and can match or exceed exact dense attention, achieving up to a 5.30× speedup at 512k tokens and significant gains on retrieval-heavy tasks.
UniPrefill is a prefill‑acceleration framework that works with virtually any model architecture, directly speeding up token‑level computation. It is implemented as a continuous‑batching operator and extends vLLM’s scheduling to support prefill‑decode co‑processing and tensor parallelism. The method delivers up to a 2.1× speedup in Time‑to‑First‑Token, with gains growing as concurrent requests increase.