MASS: One Model, Three Chairs — Recursive Self-Improvement via Multi-Agent Self-Supervision

> English summary of a Chinese zhichai.net article introducing MASS (Multi-Agent Self-Supervision), a recursive self-improvement (RSI) framework from UC Berkeley and Sakana AI (arXiv:2610.12176). The source is a long-form essay; below is a structured summary preserving its key arguments and numbers.

Key points

• The setup. The article opens with a metaphor: a late-night office with three chairs — the worker's desk, the reviewer's seat, and the consultant's seat. MASS puts the same base model (Qwen3.6-27B) in all three roles:
• Agent: executes tasks according to the current workflow.
• Evaluator: performs pairwise comparisons between outputs (comparisons rather than absolute scores — easier to get right and harder to game).
• Optimizer: reads evaluations plus full history and rewrites the workflow itself.
• Why this matters: the supervision bottleneck. As AI produces open-ended research outputs, human experts will eventually be unable to judge quality. But a single model judging its own output suffers from a "homogeneous loop" failure — shared priors reinforce errors rather than correct them. MASS's answer: don't let one instance be one judge — split it into three. Drawing on Minsky's *Society of Mind*, intelligence emerges from organized interaction of simple components; here, structure itself becomes the error-correction mechanism.
• Workflows are just text. A workflow is a four-part specification of an orchestrator–subagent architecture:
1. role — who is on the team; 2. instruction — each member's job description; 3. contract — what information subagents must return; 4. hop — the call/ordering sequence.
• Inner loop: evolve the blueprint, don't touch the weights. Evolutionary search with elite size 1: the Agent produces, the Evaluator writes comparison verdicts, the full history enters context, and the Optimizer writes the next workflow version. Notably, the Optimizer mainly edits contract and hop (information structure), rarely role and instruction — mirroring a veteran manager's instinct that efficiency comes from how information flows, not from personnel changes.
• Outer loop: workflow internalization. The champion workflow generates 18 trajectories, ranked via a pairwise tournament converted with a Bradley-Terry model; top 16 are kept (15 train / 1 validation). Training uses LoRA (rank 64), 119 trajectories, 2,435 training windows, 17.2M supervision tokens, with a 2:1 subagent-to-orchestrator sampling ratio. Crucially, the workflow text is deleted from the prompt during training — the model must internalize the organizational skill rather than read it from a script. Two rounds stack: L(0) → L(1) → L(2).

Results

• Score per output token on four research benchmarks (MLR-Bench, ScienceAgentBench, DSBench, AstaBench): 1.2–1.6x the base model.
• Unseen test tasks (3 held out from 12 synthetic GPT-5.5-generated research tasks spanning finance, robotics, pharmacology): win rate rises 53.9% → 69.9% — gains transfer, so it's not memorization.
• Evaluator accuracy improves 73% → 93%, so the improved model becomes a better judge for the next round — the recursive flywheel turns. All three roles improve together.
• Control experiment: students trained on multi-agent trajectories beat those trained on single-agent trajectories (68.3% vs 64.0% win rate), and the single-agent group used 1.4x more training tokens. Collective coordination skill can be distilled into shared weights.

Ablation: the bottleneck is the optimizer, not the evaluator

Counter to intuition, upgrading the evaluator alone does not unlock the loop — workflow optimization is the bottleneck of homogeneous RSI. The scarce capability is rewriting the organization itself, which explains why the optimizer concentrates on contracts and hops.

Information relocation

A final analysis figure shows a clean division: after optimization, role and instruction become more generic (task-specific information rinsed out), while contract and hop become more task-specialized — information ends up stored exactly where it belongs, a concrete realization of Minsky's organized interaction.

Caveats

The evaluation uses synthetic tasks generated by GPT-5.5; the base is a 27B model with LoRA-level adaptation; benchmarks are research-specific. The supervision bottleneck is loosened, not solved — a model judging itself still operates within rules it defined.

References

1. MASS: Recursive Self-Improvement through Multi-Agent Self-Supervision. Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi, Matei Zaharia, Yujin Tang. UC Berkeley & Sakana AI, 2026-10-08. arXiv:2610.12176. https://arxiv.org/abs/2610.12176 2. Recursive Harness Self-Improvement. Lee et al. arXiv:2607.15524. https://arxiv.org/abs/2607.15524 3. Marvin Minsky. *Society of Mind*. Simon & Schuster, 1986.