The Pulse
Sakana AI Tests a Two-Cycle Loop for Agent Self-Improvement
Sakana AI describes MASS, a system that uses a shared model to coordinate virtual agents, evaluate their work and train on selected runs. After two cycles, the model improved its score per output token on four research benchmarks, while res
AI.info Team · Written with AI ·
In October 2026, Sakana AI published results for a system that asks one language model to improve how it works with a team of virtual agents, then learn from the team’s strongest runs. The method, called Multi-Agent Self-Supervision (MASS), completed two training cycles with Qwen3.6-27B. Across four research benchmarks, the resulting model scored 1.2 to 1.6 times the base model’s performance per output token, the company said.
The result is a small-scale test of recursive self-improvement, not a model rewriting its own code or training an entirely new successor. MASS changes the agents’ workflow, selects executions using the model’s own judgments, and fine-tunes the shared model on those executions. The research paper was submitted to arXiv on October 8, 2026.
One model fills three roles
MASS assigns the same model three jobs: solving a task, proposing how a team of agents should divide and coordinate the work, and judging which results are better. In the workflow search, agents execute a proposed plan; an evaluator compares the results with the current best; and an optimizer uses that feedback to revise the plan. The model then ranks completed runs and trains on selected agent conversations and tool use.
The study tested the method on 12 synthetic research tasks involving finance, robotics and pharmacy. Nine tasks were designated for training and three were held out; one training task did not yield a workflow that beat its reference, leaving eight tasks to supply training data. The authors also say the task descriptions were generated by GPT-5.5. The experiment therefore tests a defined research-task setting, not unrestricted self-directed scientific discovery.
“However, a single model instance struggles to critique and improve its own complex reasoning under this homogeneous loop.”
Hyunin Lee et al., authors of “Recursive Self-Improvement through Multi-Agent Self-Supervision,” write
Research gains, uneven transfer
On three synthetic test tasks, external judges preferred the model after two MASS cycles over the base model in 69.9% of comparisons, up from 53.9% after one cycle. For public research benchmarks, Sakana AI reports score-per-output-token ratios of 1.60 on MLR-Bench, 1.35 on DSBench, 1.20 on ScienceAgentBench and 1.28 on AstaBench E2E-Hard. The paper says GPT-5.5 and Claude Opus 4.8 provided these outside assessments, and their judgments did not guide MASS’s training loop.
Improvement did not carry over to every test. The second-round model reached 0.96 times the base model’s score per output token on Terminal-Bench 2.0 and 0.94 times on SWE-bench Verified. Those results put a boundary around the headline gains: the strongest improvements appear on research tasks closer to the system’s training setup, while software benchmarks show weaker transfer.
The evaluator is not the only bottleneck
The paper also examines where the loop can stall. In a workflow-search comparison, replacing only the evaluator with GPT-5.5 helped less than replacing both the evaluator and optimizer; the authors interpret that result as evidence that an optimizer may struggle to act on feedback even when the feedback itself is stronger. The fully homogeneous loop eventually found winning workflows for 11 of 12 tasks, but took longer.
Sakana AI warns that self-judgment creates a safety problem: if the model’s evaluator shares a blind spot with the model being improved, it may accept a flawed result as progress and reinforce it through training. The reported gains are benchmark results from two cycles, not evidence that the method can improve indefinitely or reliably assess open-ended discoveries. The paper’s clearest limitation is its own benchmark split: 1.2 to 1.6 times the baseline on four research measures, but below baseline on the two software measures.