{"event":{"id":"evt-rts-20261011-mass-two-round-rsi-preprint","dedupe_key":"research_method_experimental_result:2026-10-08:mass-researchers-report-two-bounded-recursive-multi-agent-self-improvement-cycles","event_type":"research_method_experimental_result","title":"MASS researchers report two bounded recursive multi-agent self-improvement cycles","summary":"UC Berkeley/Sakana AI researchers published MASS on October 8, 2026. A Qwen3.6-27B model was used as workflow optimizer, evaluator and executor across two laboratory cycles with model weight updates based on self-selected multi-agent trajectories. Reported external-LLM-judge win rates on three held-out synthetic tasks versus the base model were 53.9% after cycle 1 and 69.9% after cycle 2, with cycle 2 winning 60.8% against cycle 1. Research benchmarks reached 1.2–1.6x base score per output token; Terminal-Bench 2.0 (0.96x) and SWE-bench Verified (0.94x) did not improve. This reports bounded preprint evidence, not independently verified indefinite or accelerating recursive self-improvement, a new civilization transition, or proven deployment.","occurred_at":"2026-10-08T15:44:02.000Z","published_at":"2026-10-08T15:44:02.000Z","observed_at":"2026-10-11T18:29:24.000Z","reconstructed_at":null,"ingest_type":"live_evidence_only","locations":null,"status":"reported","confidence":0.92,"metadata":{"source_scope":"researcher preprint","epistemic_boundary":"Two cycles only; external LLM judges; small synthetic held-out task suite; uneven transfer to software engineering; compute-normalized total costs and independent replication not established.","supported_inference":"bounded model-and-workflow feedback demonstrated in a reported research setting","unsupported":["sustained recursive improvement beyond two cycles","indefinite intelligence explosion","frontier autonomous AI research","new civilization transition","new structural relation onset"]},"created_at":"2026-10-11T20:05:23.450Z","updated_at":"2026-10-11T20:05:23.450Z"},"artifacts":[{"id":"art-mass-rsi-paper-261012176","dedupe_key":"3c938a3da4d451bf68f81888832958c80b9b0d0dddc5f87984b24cc3d5c52869","canonical_url":"https://arxiv.org/abs/2610.12176","evidence_locator":"","artifact_type":"research_preprint","title":"Recursive Self-Improvement through Multi-Agent Self-Supervision","summary":"October 8, 2026 preprint by UC Berkeley and Sakana AI researchers. Two bounded recursive iterations of workflow optimization and fine-tuning with Qwen3.6-27B occupying executor, optimizer and evaluator roles.","creator_entities":["Hyunin Lee","Jinglue Xu","Jeffrey Seely","Donghyun Lee","Somayeh Sojoudi","Matei Zaharia","Yujin Tang"],"released_at":"2026-10-08T15:44:02.000Z","content_hash":null,"metadata":{"evidence_locator":"arXiv:2610.12176 v1 (39 pages)","uncertainty":"Preprint, author-reported experimental results and LLM-judge evaluation; no independent replication."},"created_at":"2026-10-11T20:05:23.448Z","updated_at":"2026-10-11T20:05:23.448Z","relation":"evidenced_by"}],"signals":[]}