A new paper proposes letting an AI model retrain itself by splitting into a team of copies that critique and refine each other's work.
Researchers describe a method called Multi-Agent Self-Supervision, or MASS, that alternates between evolutionary search over multi-agent workflows and fine-tuning on the results. A single base model, reported as a roughly 27-billion-parameter Qwen model, proposes, runs, and grades different arrangements of itself acting as multiple agents with distinct roles; the best-performing arrangements are then folded back into training through supervised fine-tuning. Over two rounds of this cycle, the approach produced 1.2 to 1.6 times better performance per output token across four open-ended benchmarks. A separate model trained on the resulting multi-agent transcripts outperformed a single-agent model trained on 1.4 times more tokens.
The appeal here is dodging the usual ceiling on self-improvement research: once a model's reasoning gets good enough, there is no human expert left who can reliably grade it, so the model has to grade itself. Turning one model into a panel of critics that score and refine each other's work is one way around that, and the token-efficiency gains suggest this is more than a benchmark trick.
Recursive self-improvement claims have a history of overpromising; two cycles across four benchmarks is a promising result, not proof that a model can rewrite its own soul unsupervised.