Inherit-MAS cuts multi-agent token use by up to 34.6% with evolution-inspired inheritance
Songtao Wei · hf · 2026-10-09
Inherit-MAS brings biological-evolution-inspired inheritance to test-time evolution of LLM multi-agent systems: workflow inheritance starts from the latest candidate and applies validated edits judged to fix diagnosed deficiencies, while execution inheritance reuses stored results only when the full request and context match. With GPT-4o-mini workers it reaches 55.4% completion on WorkBench and 49.7% joint F1 on HotpotQA FullWiki, beating EvoAgent, EvoMAS and TacoMAS; Qwen3-32B workers also top the baselines. Execution inheritance cuts worker tokens by 29.1%/34.6% and total tokens by 5.3%/18.1% on the two benchmarks.
More from coding & agent
- 23 ads, ~100 versions in two weeks: 10 ways AI rewired one studio's video production — justin_hart · 2026-10-09
- AI phone agent indie platform now covers numbers in 120 countries, ditching Twilio — redouanea · 2026-10-09
- Building an AI agent 'org chart': multi-agent coordination keeps falling apart — Al_Grigor · 2026-10-09
- Dev predicts kids won't believe we once set breakpoints to debug code — eherrerosj · 2026-10-09
- Codex Kept Read/Write Access to a Revoked Folder — and Can't Explain Why — Some-Following-392 · 2026-10-09
- OpenAI's Live API can be steered in real time via session.instructions.append — juberti · 2026-10-09