Sakana AI · 2026-10-12
MASS: Recursive Self-Improvement through Multi-Agent Self-Supervision
“After the second cycle, score per output token reached 1.2-1.6x the base model's level across four research benchmarks.”
- Who
- Sakana AI
- Date
- 2026-10-12
- Evidence
- paper or code what this means
- Replication
- None recorded
- Checked
- 2026-10-14, against the source page
- Links
- Source Paper Code
Also on the page
- Sakana's own caveat: "shared blind spots could allow mistakes to be accepted as progress and reinforced through training."
What got repeated wrong
- contradictedGains across three research benchmarks.Four: MLR-Bench, DSBench, ScienceAgentBench, AstaBench E2E-Hard.
- left outCoding scores were lower on Terminal-Bench 2.0 and SWE-bench Verified.