Intel / article

Recursive Self-Improvement through Multi-Agent Self-Supervision

Source: unknown
Author: Sakana AI
Date: Published Oct 11, 2026

Why it matters

Reports that training on the best multi-agent team trajectories improves both solver and evaluator roles, and that multi-agent traces are more efficient training data than single-agent ones.

Key takeaways · AI-distilled

Sakana AI's MASS has one shared model act as a team of virtual subagents, search for better workflows, judge which work best, then train on the best team's executions to distill that collective experience back into the model.
Over two cycles with a 27B open-weights model on synthetic open-ended research tasks, score per output reached 1.2 to 1.6x the base model's level across four research benchmarks, the authors report.
Training only on task-solving data also improved the model's roles as optimizer and evaluator, and multi-agent trajectories proved more efficient training data than single- ones. The authors found optimizer capacity, not evaluator capacity, could limit improvement speed.
The team flags a safety risk in this homogeneous loop: when a model is its own evaluator, shared blind spots could let mistakes be accepted as progress and reinforced through training, so detection and intervention methods need study.

Terms in this piece · Glossary

multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Read the source unknown