AI Wiki · Mediocre Security
CompanyModels & inference infrastructure

Sakana AI

Japan-based lab.

Japan-based lab. MASS (Oct 12 2026) is a recursive self-improvement framework where one 27B open-weights model acts as executor, workflow optimizer and evaluator, reporting 1.2-1.6x score per output token, with the paper flagging shared blind spots when a model grades itself (Sakana). Multi-Layered Review is a three-agent Claude-based paper reviewer that caught 73.43% of planted core-claim errors but 16.11% on real withdrawn arXiv papers (TechAIMag).

MASS in detail

MASS was published with UC Berkeley researchers on Oct 11 2026 (arXiv 2610.12176). One open-weights model, Qwen3.6-27B, plays solver, workflow optimizer and evaluator, and the optimized multi-agent workflows become the self-supervision signal for the next round. Sakana reports two cycles, with round two scoring 1.2-1.6x the base model per output token; GPT-5.5 and Claude Opus 4.8 served as outside judges. Gains concentrate on tasks resembling the training data, and two software-engineering benchmarks dipped slightly. The authors call a shared blind spot in a homogeneous loop the “Achilles’ heel” of recursive self-improvement: a flawed result can be produced, accepted as an improvement and baked into the next model. It is a small-scale version of the automated-research capability Recursive markets and that Anthropic’s IPO filing flags as an approaching threshold. Figures are self-reported.

Timeline

Every dated event in the wiki that involves Sakana AI, newest first.

  • Research
    Sakana AI publishes MASS, a recursive self-improvement framework

    With UC Berkeley: one Qwen3.6-27B model acts as solver, optimizer and evaluator; round two scores 1.2-1.6x the base per output token (self-reported). Authors warn shared blind spots can compound across rounds.

Connections

Sakana AI and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →

Sources

  1. AI Weekly, “Sakana AI publishes MASS”, Oct 2026
  2. remio, “Sakana AI reports recursive self-improvement gains, but oversight remains a constraint”