Skip to main content
Back to timeline
BlogSource publication:

A model coaches and judges itself: two rounds of multi-agent self-supervision lift score per output token to 1.2–1.6x baseline

Synopsis

One shared model plays a team of virtual subagents, searches for better multi-agent workflows, picks the best by its own judgment, and trains on that team's executions; this is Multi-Agent Self-Supervision (MASS). With a 27B open-weights model over two cycles, score per output token reached 1.2–1.6x the base model's level on four open-ended research benchmarks. The work also warns that when a model is its own evaluator, shared blind spots could let mistakes be accepted as progress and reinforced through training.

Source-provided article image: Recursive Self-Improvement through Collective Intelligence
sakana.ai

Interpretation

A single model can solve open-ended research tasks through a team of virtual subagents and learn from the team's best workflows, without external experts judging each result. In earlier self-improvement loops the model was both optimizer and evaluator, lacking useful self-supervision and risking reinforcement of its own blind spots rather than correction. Two cycles with a 27B open-weights model on synthetic open-ended research tasks; after the second cycle, score per output token reached 1.2–1.6x the base model's level across four research benchmarks.

Training data comes from team executions, and the team's collective experience is distilled back into the shared model, which gains both optimizer and evaluator capability. MASS has one shared model solve tasks through virtual subagents, search for better multi-agent workflows, and select the best ones using its own judgments. Training only on task-solving data from the optimizee improved both optimizer and evaluator, indicating role generalization within the loop; multi-agent trajectories proved more efficient training data than single-agent ones.

Perspective

The reported scope is one 27B open-weights model, synthetic open-ended research tasks, two self-improvement cycles, and score per output token on four research benchmarks. It is aimed at researchers and engineers working on self-improvement training loops and multi-agent system safety. The source states no next experimental step.

The source frames homogeneous self-improvement as a safety question the community needs to investigate carefully: when a model is also its own evaluator, shared blind spots could allow mistakes to be accepted as progress and reinforced through training. It also reports that optimizer capacity, rather than evaluator capacity, can be the bottleneck in improvement speed, a judgment drawn from the two reported cycles.