Sakana AI's MASS Scales Recursive Self-Improvement via Multi-Agent Self-Supervision
omarsar0 · x · 2026-10-10
Sakana AI proposes MASS, a way to scale recursive self-improvement without external verifiers.
- Standard RSI loops need an external checker, leaving open-ended tasks out; MASS removes that requirement
- One base model proposes multi-agent workflows, runs and grades them, and evolutionary search keeps the best scorers; the model is fine-tuned on its own traces and returns as a better optimizer and grader
- Two cycles on Qwen3.6-27B raise performance per output token from 1.2x to 1.6x on four open-ended benchmarks
- A student trained on multi-agent traces beats a single-agent student trained on 1.4x more tokens
More from Research
- PopVax's AI-designed fridge-stable RNA vaccine enters Phase I human trial — SeyoneC · 2026-10-11
- Stanford paper simulates AI eval ecosystem with LLM agents: holdout design tradeoffs — sanmikoyejo · 2026-10-11
- First AI Measurement Science workshop at COLM 2026 targets rigorous AI evaluation — sanmikoyejo · 2026-10-11
- Nat Lambert advises Mercor on RL data: evals science is the next frontier — natolambert · 2026-10-11
- Open-source LLM simulator separates structural effects from LLM quirks in benchmark markets — sanmikoyejo · 2026-10-11
- Simulation platform tests how evaluator independence and mandatory disclosure reshape AI markets — sanmikoyejo · 2026-10-11