MASS

Overview

MASS is a Sakana AI self-improvement method (per The Decoder, Oct 12) that puts a Qwen3.6-27B into three roles — solver, judge, refiner — with LoRA on best runs. The paper’s own caveat is that shared blind spots risk being reinforced across self-training rounds rather than caught — a documented limitation, not a design feature. MASS extends Sakana AI‘s broader non-frontier-capex RSI thesis documented in the June 2026 Tokyo RSI Lab announcement.

Timeline

2026-10-13-AI-Digest — MASS surfaces in The Decoder’s Oct 12 coverage as one of four agent-adjacent findings landing in a 24–48 hour window (alongside Fork-and-Flush, X-AgentRed, and RATED). Three roles — solver, judge, refiner — on a shared Qwen3.6-27B with LoRA on best runs. Load-bearing softener: the shared-blind-spots risk is Sakana's own caveat, not a critique layered on top — the paper is explicit the limitation is structural. Reframe worth carrying: MASS extends Sakana's non-frontier-capex RSI bet into a three-role self-training recipe with a surfaced blind-spot limitation, not MASS catches its own blind spots. Pairs this week with Fork-and-Flush as two agent-architecture signals landing on The Decoder in the same 48-hour window.

Key Developments

Three-Role Self-Training Recipe (Solver / Judge / Refiner) on Qwen3.6-27B With Documented Shared-Blind-Spot Risk (October 13, 2026): MASS puts a Qwen3.6-27B into three roles — solver, judge, refiner — with LoRA applied on best runs. The headline load-bearing caveat is in the paper itself: shared blind spots across the three roles risk being reinforced across self-training rounds rather than caught. The method extends Sakana AI‘s non-frontier-capex RSI thesis from a single-model-rewrites-itself frame to a role-split self-training frame. The honest framing to carry: solver/judge/refiner is a decomposition, not a diversification — shared training lineage limits the diversity of the critique channel.

MASS

Overview

Timeline

Key Developments

Related