Sakana AI's MASS: multi-agent self-supervision yields 1.2-1.6x per-token gains for recursive self-improvement
hardmaru · x · 2026-10-12
Sakana AI, with UC Berkeley, introduces MASS (Multi-Agent Self-Supervision), targeting the "supervision bottleneck" in recursive self-improvement: when tasks outstrip human evaluation, self-evaluation risks reinforcing a model's own blind spots. Inspired by Minsky's Society of Mind, MASS has a single base model organize copies of itself into a structured multi-agent team, alternating between evolving its own multi-agent workflows (roles and information routing) and fine-tuning on the team's generated trajectories. Over two MASS cycles on Qwen3.6-27B, the model achieves 1.2–1.6x higher performance per output token across four public benchmarks, and a student trained on multi-agent traces outperforms a single-agent student trained on 1.4x more tokens — suggesting joint orchestration learning can be an effective RSI signal.
Related event: Sakana AI Unveils MASS for Recursive Self-Improvement(3 posts)→
More from coding & agent
- David Fowler: build CLIs for agents to explore systems and fix bugs autonomously — mariorod1 · 2026-10-12
- rtoml v0.14 Adds Python 3.15 Support, i128 Ints, and Big Speedups — samuelcolvin · 2026-10-12
- Reverse-engineering ChatGPT's Intelligent UI: DIL language, server-side compilation and sandboxed runtime — rohanpaul_ai · 2026-10-12
- Reverse engineering leaks three unannounced Grok Bot features, versioned memory leading — bfrench · 2026-10-12
- badlogic demos durable codemode: persisting agent state so long tasks survive aborts — max_paperclips · 2026-10-12
- supermemory packages agent context into one API with memory, retrieval built in — julianweisser · 2026-10-12