A model grading its own work drifts: give it a virtual team and score per output token rises to 1.2-1.6x baseline after two cycles

The model no longer solves and grades tasks on its own; it first assembles a team of virtual subagents, lets that team do the work, and learns from those executions. In the earlier self-supervision loop the model was both optimizer and evaluator inside a homogeneous loop, where it risked reinforcing its own blind spots instead of correcting them. Two cycles were run 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.

The recipe has one shared model act as several virtual subagents, search for better multi-agent workflows, and then use its own judgments to select the best ones. Multi-Agent Self-Supervision (MASS) takes the selected best team's executions as training data and distills the team's collective experience back into the shared model. Training only on task-solving data from the optimizee still improved both the optimizer and the evaluator, indicating role generalization within the loop.

The demonstrated scope is one 27B open-weights model, two cycles, synthetic open-ended research tasks, and score per output token on four research benchmarks. It speaks to researchers and engineers concerned with recursive self-improvement and multi-agent training, and asks whether self-supervision in a homogeneous loop yields effective gains. The authors state that the community needs to investigate homogeneous recursive self-improvement carefully to support AI safety, which is the next direction the material explicitly proposes.

The authors note that when a model is also its own evaluator, shared blind spots could allow mistakes to be accepted as progress and reinforced through training. They argue that for AI to improve beyond human expertise safely, these failures need to be understood and effective ways to detect and intervene preserved. These are open questions the authors raise; the material does not present evidence that they are resolved.