One model teaches its own team
Sakana AI published the paper Recursive Self-Improvement through Multi-Agent Self-Supervision. The method is Multi-Agent Self-Supervision, MASS. The note links the paper and the code.
The question is whether a model can learn from the collective judgment of its own team. One shared model solves tasks through a team of virtual subagents, searches for better multi-agent workflows, and uses its own judgments to select the best ones. Training on the best team's executions distills that team's experience back into the shared model.
[1]
After two cycles, the score per token rose
The runs used a 27-billion-parameter open-weights model on synthetic open-ended research tasks, for two cycles. After the second cycle, score per output token reached 1.2 to 1.6 times the base model's level across four research benchmarks.
The note does not name those four benchmarks. The multiple is score divided by output tokens. It is not a claim that the model is stronger by that factor on every research task.
[1]Three results in the note
First, training only on task-solving data from the model being optimized also improves the optimizer and the evaluator. The note calls this role generalization. Second, for this kind of recursive self-improvement, multi-agent trajectories are more efficient training data than single-agent trajectories. Third, the bottleneck in how fast the system improves can be the optimizer's capacity, rather than the evaluator's.
[1]When the model grades itself, blind spots stay
Sakana writes that this homogeneous recursive self-improvement needs careful study, for safety. When a model is also its own evaluator, shared blind spots could let mistakes be accepted as progress and then reinforced by training. The authors say that if a system is to improve past human expertise safely, those failures have to be understood, and effective ways to detect them and intervene have to remain.
The note does not say this blind spot already appeared in the two cycles. It is a risk they mark, not a result of the score.
The authors are Hyunin Lee, Jinglue Xu, Jeffrey Seely, Yujin Tang, and, from UC Berkeley, Somayeh Sojoudi, Matei Zaharia, and Donghyun Lee.
[1]