Paper breakdown
November 4, 2026 arXiv paper: SpecCompact made Hermes-style agent context compaction 2.04–5.63× faster on AppWorld and τ³, with an 8.68× median speedup in a Hermes user study.
Paper breakdown
A November 3, 2026 technical report: compaction kept Claude Opus 4.5 and GPT 5.6 Sol near full-history accuracy while the live context stayed short.
Paper breakdown
MIT CSAIL note, November 2, 2026: GPT-6 Astra edited neural weights, with no training loop, and raised a small model's MATH-500 score from 7.9% to 48.1%.
Paper breakdown
Notes on Yao et al. 2022. ReAct interleaves a written reasoning trace with an action, then reads the observation, on HotpotQA, FEVER, ALFWorld, and WebShop.