Thesis
MASS makes the team’s accumulated work a self-supervision signal for a shared model.
Observed facts
The source says one shared language model serves as task solver, workflow optimizer, and evaluator.
1The source describes an inner workflow-search loop with fixed weights and an outer loop that trains on selected workflow executions.
1The source reports improved score per output token over the base model on research benchmarks and says external judgments were withheld from MASS.
1
Mechanism
An inner loop has the model revise and judge multi-agent workflows; an outer loop distills selected executions into the shared model’s weights, which then resumes all three roles.
Why now
The source motivates homogeneous recursive self-improvement for settings where expert assessment is scarce or slow, and reports benchmark gains from the proposed method.
Implications
Workflow search and execution traces can become training material rather than disposable orchestration.
The usefulness of the loop depends on whether its self-evaluator selects work that tracks real task quality.
Evidence boundary
The training tasks are non-verifiable and the self-evaluation signal’s correctness is unknown; external GPT-5.5 and Claude Opus 4.8 judgments were withheld from MASS and do not establish general reliability.
Watch conditions
Whether gains persist on independently verified tasks or under external evaluation across repeated runs.