LLM Digest
Subscribe

AI Storyline

3 items · 2 sources · 3 days

View as JSON

Operational story trace

Exact Language

Latest change

FrontierReplay (Nov 19) proposes exact prefix-cache reuse for hybrid attention/recurrent models, whose in-place state updates normally limit reuse to saved checkpoint positions.

Earlier contextThe story so far

This is a loose grouping, not one event: three unrelated arXiv papers that share a theme of exactness guarantees for language models. It began with SpecCTMC (Nov 5), exact speculative sampling for diffusion LMs, then a study of small models trained under an exact verifier (Nov 16).

editor-curated · source-linked

Arc

Nov 5Nov 19 · now
SAMPLING · Nov 5
SpecCTMC: parallel diffusion-LM decoding without changing the output distribution
1 source · show source ▾
VERIFICATION · Nov 16
Small models under an exact verifier learn procedure, not judgment
1 source · show source ▾
CACHING · Nov 19
FrontierReplay: split-free exact prefix caching for hybrid models
1 source · show source ▾

What to watch — open questions

  • Does FrontierReplay ship in an open-source inference server such as vLLM or SGLang?
  • What wall-clock speedup does SpecCTMC report over standard masked-diffusion decoding?
  • Do exact-verifier results hold for models larger than the small ones studied?
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.