MASS paper: two cycles lift Qwen3.6-27B per-token performance 1.2–1.6x via multi-agent self-supervision
TheTuringPost · x · 2026-10-13
The arXiv paper (Sakana AI + UC Berkeley, including Matei Zaharia) details MASS: to overcome the supervision bottleneck of recursive self-improvement on non-verifiable tasks, a single base model iteratively proposes, executes, and self-evaluates multi-agent workflows, using evolutionary search with structural guardrails to discover effective roles and information routing, followed by supervised fine-tuning — enabling continuous recursive bootstrapping.
On Qwen3.6-27B, two MASS cycles yield 1.2–1.6x higher performance per output token on four open-ended benchmarks. Multi-agent traces are also more training-efficient: a student trained on them outperforms one trained on single-agent traces with 1.4x more training tokens. Paper and code are open.
Related event: Sakana AI and Berkeley Propose MASS for Recursive Self-Improvement(4 posts)→
More from AGI Musings
- Google RCT: AI boosts patent lawyers' work quality, but juniors gain no judgment — rseroter · 2026-10-13
- Is China's commercialization lead limited to physical AI and renewables? — PAstynome · 2026-10-13
- Two camps in SWE: will superintelligent AI write machine code or JavaScript? — wordgrammer · 2026-10-13
- "You have a magic coding robot and your big idea is cloning mediocre apps?" — rms80 · 2026-10-13
- Economist Josh Gans: Writing PowerPoint slides is thinking — joshgans · 2026-10-13
- Unitree CEO: GPT-6 works "very well" on robots, GPT-7 may unlock mass deployment — 141_1337 · 2026-10-13