Reproduce “SpecCompact: Accelerating LLM Agent Context Compaction with Slot-Parallel Semantic Speculative Verification” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yulu An, Yifei Feng, Xiangyang Yin

SpecCompact shortens agent histories by drafting six summary slots in parallel with a small model, while a target model checks and repairs each line. A judge evaluation found independent slots retained relevant information at least as well as monolithic summaries. On Gemma4-31B-IT’s AppWorld Normal workload, compaction was 4.71× faster than Hermes, but task performance varied and was not always preserved.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.