LLM Digest
Subscribe

AI Storyline

4 items · 4 sources · 4 days

View as JSON

Operational story trace

Attention Sparse

Latest change

DTS (Nov 19) scores attention tiles by deltas to decide which blocks to skip, targeting the prefill bottleneck of long documents and codebases.

Earlier contextThe story so far

This is a loose cluster of four independent arXiv papers, not one event: each attacks the cost of long-context attention by letting a query attend only to a selected subset of keys. It opened with work on DeepSeek Sparse Attention's learned indexer, then two training-free methods that choose keys by semantic blocks or per-head structure.

editor-curated · source-linked

Arc

Nov 2Nov 19 · now
LEARNED INDEXER · Nov 2
Overlapping selection and attention in DeepSeek Sparse Attention
1 source · show source ▾
TRAINING-FREE · Nov 3–6
Two methods pick keys without retraining the model
2 sources · show sources ▾
LATEST · Nov 19
DTS: delta-based tile scoring for block-sparse prefill
1 source · show source ▾

What to watch — open questions

  • Do any of these methods ship in vLLM, SGLang, or TensorRT-LLM kernels?
  • Which method holds accuracy on long-context benchmarks like RULER at matched sparsity?
  • Do the training-free methods compose with KV-cache quantization and prefix caching?
How this thread was built
editor wrote the arc · 3 beats

Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.