Sakana AI's MASS makes one model self-improve by playing all three agent roles, boosting token efficiency 1.2-1.6x
量子位 · wechat · 2026-10-18
Sakana AI (co-founded by Transformer author Llion Jones) and UC Berkeley propose MASS, a recursive self-improvement method where a single model plays executor, evaluator and optimizer roles. Inner loops search text-defined multi-agent workflows while outer loops SFT the base model on golden trajectories, iterating twice.
Results: token efficiency on open-ended research tasks improved 1.2-1.6x, self-evaluation accuracy rose from 73% to 93%, and multi-agent trajectories beat single-agent data (68.3% vs 64.0% win rate) with 31% fewer tokens. Limitations: only two iterations tested, no gains on SWE-bench Verified, and workflow search is fragile.
More from Research
- Eli Ben-Sasson: wait 3-5 years before trusting lattice crypto against AI attacks — examachine · 2026-10-18
- MIT lets a Claude-powered agent freely "play" on a simulated island for 30 hours — and finds play expands its capabilities — LuizaJarovsky · 2026-10-18
- New paper measures model capability via how models solve tasks, not success rates — maksym_andr · 2026-10-18
- Economy of Minds: local incentives may replace central managers in multi-agent AI — pliang279 · 2026-10-18
- Ropedia Super Week: world model W0, dexterous bimanual controller DONIC and more — liuziwei7 · 2026-10-18
- Google Research builds AI-agent-controlled wet lab with autonomous molecular design loop — Puzzleheaded-King584 · 2026-10-18