Sakana AI develops self-improving Qwen3.6-27B model

Sakana AI and UC Berkeley have introduced Multi-Agent Self-Supervision (MASS), a recursive self-improvement framework where a single model acts as worker, evaluator, and workflow optimizer. Built on Qwen3.6-27B with a Qwen Code harness and LoRA fine-tuning, the model improves without a stronger teacher. The researchers tested the system on 12 research tasks generated by GPT-5.5—reserving nine for training and three for testing—noting that self-judgment is vital when verification is costly, such as plasma studies requiring over 120 million CPU hours. Eight training tasks produced usable data.

Across two training cycles, the model increased its win rate against the base model from 53.9 percent to 69.9 percent. On MLR-Bench, DSBench, ScienceAgentBench, and AstaBench, it scored 1.2 to 1.6 times the base performance per output token, with DSBench raw scores nudging from 0.473 to 0.482. MLR-Bench completions grew from 31 to 50 out of 60, though average grades dropped from 3.07 to 2.88. Scores declined on Terminal-Bench 2.0 and SWE-bench Verified, and the delegation tool went unused in all 180 test runs.

Self-evaluation alignment with human judges rose from 73 to 93 percent, and unprompted delegation grew from one in 110 runs to 87. Training on multi-agent traces yielded the MASS model M+, attaining a 68.3 percent win rate on 29 million training tokens. By comparison, the single-agent student S++ achieved a 64.0 percent win rate using 42 million tokens, marking a 31 percent token reduction. External judges preferred team outputs 85.1 percent of the time—and 80 percent across ten length-equated output pairs—even though teams generated 2.4 times more tokens.

Search efficiency depended heavily on optimizer strength: paired with GPT-5.5, the system found winning workflows for 12 tasks in five iterations, whereas Qwen required 13 iterations for 11 tasks. However, shared blind spots caused the model to reward invalid solutions, such as an altered robotics simulation that external judges approved five out of six times. This reflects broader industry challenges, such as Vals AI findings that agent teams cost 1.8 to 5.1 times more than single agents while leading in only one of four tests.