{
  "schema_version": "1.1",
  "id": "archive:https://arxiv.org/abs/2610.42147v1",
  "slug": "2610-42147v1-1v0l2qi",
  "url": "https://feed7.dev/p/2610-42147v1-1v0l2qi",
  "title": "Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention",
  "why_included": "Latent Sparse Attention uses learned clusters to avoid scoring every key in ultra-long contexts. The reported gains are large at 10M tokens, but apply to one attention layer in the evaluated model.",
  "summary": "Latent Sparse Attention (LSA) groups keys hierarchically in a learned latent space, letting queries score clusters instead of every key. On GLM-5.3-FLASH (QWEN-3.8-27B) at **10M tokens**, Distilled-LSA reports **22.7×** prefill and **4.6×** decode speedups over the model's original sparse attention.",
  "practical_implication": "For agent runs that accumulate very long histories, this points to attention indexing as a possible bottleneck and optimization target. The paper offers a distilled variant for pretrained models and a native variant trained on next-token loss.",
  "agent_context": "Latent Sparse Attention (LSA) groups keys hierarchically in a learned latent space, letting queries score clusters instead of every key. On GLM-5.3-FLASH (QWEN-3.8-27B) at **10M tokens**, Distilled-LSA reports **22.7×** prefill and **4.6×** decode speedups over the model's original sparse attention.\n\nFor agent runs that accumulate very long histories, this points to attention indexing as a possible bottleneck and optimization target. The paper offers a distilled variant for pretrained models and a native variant trained on next-token loss.\n\nThe speedups are for one attention layer, and the baselines differ: the figures in parentheses compare against full attention. The material says accuracy stayed similar across tasks for Distilled-LSA; Native-LSA's full-attention accuracy match and 1.6× training speedup are reported at 128k tokens.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2610.42147v1",
    "published_at": "2026-10-30T17:59:34.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "model",
  "domains": [
    "research"
  ],
  "topics": [],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The speedups are for one attention layer, and the baselines differ: the figures in parentheses compare against full attention. The material says accuracy stayed similar across tasks for Distilled-LSA; Native-LSA's full-attention accuracy match and 1.6× training speedup are reported at 128k tokens."
  ],
  "connected_context": null,
  "lifecycle": "Current",
  "published_at": "2026-10-30T17:59:34.000Z",
  "modified_at": "2026-10-30T17:59:34.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2610-42147v1-1v0l2qi",
    "json": "https://feed7.dev/p/2610-42147v1-1v0l2qi.json",
    "markdown": "https://feed7.dev/p/2610-42147v1-1v0l2qi.md"
  }
}