MASS paper: one model plays the whole agent team to bootstrap recursive self-improvement
idanbeck · x · 2026-10-28
The arXiv paper "Recursive Self-Improvement through Multi-Agent Self-Supervision" (MASS) tackles the supervision bottleneck of RSI on non-verifiable tasks. A single base model iteratively proposes, executes and self-evaluates multi-agent workflows — acting as orchestrator, workers and judge — then fine-tunes on the team's self-generated traces, with evolutionary search discovering the best roles and information routing per task. Over two MASS cycles with Qwen3.6-27B, the model gains 1.2-1.6x score per output token on four open-ended benchmarks, though an accompanying explainer notes it's flat (0.96x) on Terminal-Bench.
More from coding & agent
- Local AI Music Pipeline: LLM Subagents + YuE2 Produce a Full Eurodisco Track — Robonotes1760 · 2026-10-28
- Granting Codex or Claude sysadmin access is powerful and dangerous — RealGeneKim · 2026-10-28
- Capy moves mobile dev to the cloud with cloud-phone build and testing — k7agar · 2026-10-28
- Replit designers shipped 765 PRs and 250K lines of code in a year — amasad · 2026-10-28
- Founder: AI coding gives startups a structural edge — more hours, more risk, less coordination — yenkel · 2026-10-28
- Reddit thread: What have people actually automated end-to-end with AI agents? — IndicationTypical972 · 2026-10-28