Sakana AI says its MASS method could let AI systems improve on tasks too complex for humans to reliably judge, replacing a key training bottleneck.
Key Takeaways
Sakana AI published its MASS paper Sunday, proposing agents that supervise each other to generate training signals without external validation
MASS targets tasks too complex for humans to reliably judge, where traditional training cannot depend on labeled data or human feedback
The method uses multiple AI agents working in parallel to review and critique each other’s work
Sakana AI has not published benchmark comparisons against single-agent reinforcement learning baselines, leaving independent verification of performance gains outstanding
The paper, titled Recursive Self-Improvement through Multi-Agent Self-Supervision, or MASS, was published Sunday and shared by AI researcher hardmaru, one of the lab’s own researchers.
The core problem MASS addresses is what happens when AI models tackle tasks so complicated that no human reviewer can confidently say whether the output is right. Traditional AI training relies on labeled data or human feedback to tell a model when it has succeeded or failed.
hardmaru described the challenge directly, asking how AI can keep improving “when the tasks become too complex for humans to reliably judge.” MASS answers that by replacing a single human or single AI judge with a group of agents that supervise each other, cross-checking outputs and generating their own training signal without waiting for external validation.
What Multi-Agent Self-Improvement Actually Means
Sakana AI describes MASS as generating training signals through agents that supervise each other rather than depending on externally labeled answers.
Sakana AI’s proposed fix uses multiple AI agents working in parallel, each reviewing and critiquing the others’ work, with disagreement and consensus patterns acting as a proxy for correctness.
The approach frames the technique specifically as a recursive loop, meaning each round of agent critique directly feeds into the next generation of training.
Also Read: AI Agents Unlock 657-Fold Cost Cuts, New Benchmark Finds
A Crowded Race To Solve AI’s Scaling Ceiling
Self-improving AI has become one of the most contested research fronts in the field this year. Several major labs have separately pushed reinforcement learning techniques that turn checkable tasks into reward signals, a related but distinct approach that depends on tasks having clear right answers.
Sakana AI’s MASS paper differs by targeting the much harder case of tasks without any checkable ground truth at all, a gap that has limited how far current post-training methods can scale.
The stakes are structural. As AI labs push models toward longer, more autonomous tasks, the supply of human-labeled training data has not kept pace with the complexity of what models attempt.
If self-supervision loops like MASS can reliably substitute for human judgment on complex tasks, it removes one of the clearest bottlenecks constraining how far current training methods can scale, independent of how much compute labs can buy.
What To Watch From Here
Sakana AI has not published benchmark comparisons against single-agent reinforcement learning baselines in the material shared so far, so independent verification of MASS’s actual performance gains remains outstanding.
Researchers will be watching for a full paper release with ablation studies, the kind of experiments that isolate exactly which part of the multi-agent setup drives any improvement, before drawing firm conclusions about its real-world reliability.
Read Next: Hugging Face Clones OpenAI’s Math Tests, Admits Verifier Is ‘Early And Rough’