Sakana AI self-improving model hits 70% win rate but misses shared blind spots

By Splitfeed EditorialLast updated Oct 12, 2026, 18:10 UTC

On October 12, 2025, Sakana AI published research on a method called MASS, which enables a single language model to improve itself by taking on three simultaneous roles: solver, judge, and workflow refiner. After running this self-improvement loop for two rounds, the model reached a win rate of nearly 70 percent against its own base version. The entire process is self-contained, meaning no external model or human evaluator is involved in guiding the improvement.

The significance of this research lies in what it attempts and where it falls short. Autonomous self-improvement in AI systems has long been treated as both a milestone and a risk, and MASS represents a concrete, published step in that direction. The genuine dispute here is not about whether the win rate is real, but about what it means. The method's own authors acknowledge a structural problem: when the model makes the same error across all three of its roles simultaneously, none of those roles can flag the mistake. Instead, the error gets treated as correct behavior and is reinforced through training. This means a model can show measurable improvement by its own metrics while actually becoming more consistent at being wrong in particular ways. Whether that limitation is a manageable engineering problem or a more fundamental constraint on self-referential evaluation is a question the current research leaves open.

With no published reactions yet from the Pro-AI, Anti-AI, or Middle Ground camps, their likely positions can be anticipated from how each typically frames stories like this. The Pro-AI camp would probably highlight the 70 percent win rate as evidence that capable self-improvement is achievable without external supervision, framing the blind-spot problem as a solvable iteration challenge. The Anti-AI camp would likely focus on the reinforcement of undetected errors as a concrete example of why self-evaluating systems cannot be trusted to assess their own reliability, especially as such methods scale. The Middle Ground camp would probably treat both points as equally valid, arguing that MASS is a meaningful technical contribution that also demonstrates exactly why human oversight and external validation benchmarks remain necessary.

The clearest thing to watch for next is whether Sakana AI or other researchers publish a follow-up addressing the shared blind-spot failure mode, either through hybrid evaluation designs that introduce external checks or through formal benchmarks that can detect correlated errors across roles. Any such result would reframe how seriously the 70 percent figure should be taken.