---
id: "lib-2026-10-12-recursive-self-improvement-through-collective"
type: "library"
title: "Recursive Self-Improvement through Collective Intelligence (MASS)"
lang: "JA"
org: "Sakana AI"
source_type: "Paper"
date: "2026-10-12"
week: "2026-W42"
url: "https://nosov.ai/library/lib-2026-10-12-recursive-self-improvement-through-collective/"
original_url: "https://sakana.ai/mass/"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Sakana AI and UC Berkeley preprint (with Matei Zaharia) that states its method, compute and judging protocol, reports three-trial variance and includes the coding benchmarks where it slightly lost; not yet peer reviewed."
summary_basis: "full-text"
original_minutes: "40"
tags: ["ai technology", "recursive self improvement", "multi agent workflows"]
---

# Recursive Self-Improvement through Collective Intelligence (MASS)

Tier A (Strong evidence): Sakana AI and UC Berkeley preprint (with Matei Zaharia) that states its method, compute and judging protocol, reports three-trial variance and includes the coding benchmarks where it slightly lost; not yet peer reviewed.

Summary based on the full text.

## In short

"Recursive Self-Improvement through Multiagent Self-Supervision" (arXiv 2610.12176, 8 October 2026) by Hyunin Lee and co-authors at UC Berkeley and Sakana AI. It tackles self-improvement on open-ended research tasks, where no human or verifier can reliably grade the output and the model must act as its own solver, optimiser and judge. MASS has one base model propose, run and self-evaluate multi-agent workflows, evolves those workflows, then fine-tunes the model on the resulting orchestrator and subagent trajectories and repeats. The claim is that splitting work across copies of the model gives a better training signal than one model critiquing itself.

## Key numbers

- 1.2-1.6x: score per output token after two MASS cycles versus the base Qwen3.6-27B on four research benchmarks
- 53.9% to 69.9%: win rate of the trained model against the base model on three held-out synthetic research tasks, after one and two cycles
- 73% to 93%: self-evaluator agreement with external strong-model judges
- 68.3% vs 64.0%: win rate of a student trained on multi-agent traces versus one trained on single-agent traces with about 1.4x more tokens
- 0.96x and 0.94x: score-per-token ratio on Terminal-Bench 2.0 and SWE-bench Verified, where MASS did not help

## What it reports

- Two MASS cycles on Qwen3.6-27B raise score per output token 1.2-1.6x on MLR-Bench, ScienceAgentBench, DSBench and AstaBench.
- On the coding benchmarks Terminal-Bench 2.0 and SWE-bench Verified the ratios are 0.96x and 0.94x, a small loss.
- Win rate against the base model on three held-out synthetic research tasks rises from 53.9% after one cycle to 69.9% after two.
- Training only on task-solving trajectories also improves the other roles: self-evaluator agreement with external judges rises from 73% to 93%, and better workflows are found faster.
- Multi-agent traces train more efficiently: 68.3% win rate versus 64.0% for a single-agent student trained on about 1.4x more supervised tokens.
- Ablations point to workflow optimisation, not evaluation, as the main bottleneck in the loop.
- After training, the model delegates and hands off code even without the optimised workflow in its prompt: delegation in 87 of 110 bare-prompt runs after two cycles versus 1 of 110 for the base model.

## How they got it

Training on trajectories from eight synthetic research tasks (finance, robotics, pharmacy), testing on three held-out tasks and six public benchmarks in a fixed agent scaffold (qwen-code), three trials each. Final quality is judged by GPT-5.5 and Claude Opus 4.8 at maximum reasoning effort, three seeds each, kept out of the training loop. Fine-tuning runs on six H100 GPUs; the first-cycle SFT takes about 19 hours.

## Limits

- Preprint, one 27B base model, two cycles; whether gains keep compounding is untested and the authors name faster per-cycle gains as future work.
- Held-out task results rest on LLM judges, which carry their own biases even when kept separate.
- Gains are per output token on research benchmarks; coding benchmarks did not improve.

## Why it matters for companies

The practical lesson sits below the self-improvement headline: structuring work as a team of specialised agent roles, with explicit handoffs, produced better outputs and better training data than one long agent trace. Firms tuning their own open-weight models on internal workflows can take two points from it: multi-role traces may be worth more per token than single-agent logs, and a fine-tuned model can learn to coordinate without the workflow prompt. The result on coding benchmarks is a reminder to measure each task type separately.

## Key passages in translation

The paper is published in English (Sakana AI is Tokyo-based; no Japanese version was found), so passages are quoted as written.

> "a single model instance struggles to critique and improve its own complex reasoning under this homogeneous loop."

No translation needed: the authors' statement of the core problem.

> "Our ablations further identify workflow optimization, rather than evaluation, as a major bottleneck in the homogeneous RSI loop"

No translation needed: the main ablation finding.

Original: https://sakana.ai/mass/