Recursive self-improvement through collective intelligence

MASS (Multi-Agent Self-Supervision) proposes a homogeneous recursive self-improvement loop in which a single language model serves simultaneously as task executor, workflow optimizer, and evaluator to generate its own self-supervision from collective multi-agent workflows. It targets non‑verifiable, long‑horizon research tasks where human or rule‑based evaluation is scarce: by evolving how identical subagents coordinate and audit one another, the system elicits collective intelligence that the shared model can distill into improved weights. A workflow prompt encodes role, instruction, contract, and hop (call order), so the optimizer can focus on what information agents must produce and when; an example finance workflow enforces explicit data validation contracts to prevent common errors like future leakage.

The training loop has an inner search that iteratively executes candidate workflows, has the model self‑evaluate candidate vs. champion workspaces, and revises workflows accordingly. The outer loop runs each champion 18 times, uses the model’s pairwise preferences (converted by a Bradley-Terry model) to rank workspaces, keeps up to 16 top trajectories (15 train, 1 val), strips workflow text from the orchestrator prompt, and fine‑tunes the shared model (LoRA) on orchestrator and subagent conversations. External judges (GPT‑5.5 and Claude Opus 4.8) are withheld from training; experiments beginning from Qwen3.6‑27B show improved score per output token on research benchmarks after two MASS rounds.