Sakana AI Releases MASS Paper

Sakana AI and researchers from UC Berkeley have released the MASS (Multi-Agent Self-Supervision) paper. This method is a homogeneous recursive self-improvement (RSI) framework where a language model acts as task performer, optimizer, and evaluator without external signals, using the work trajectories of multi-agent teams composed of model replicas as fine-tuning data.

How MASS Works

MASS iterates between an inner loop (workflow exploration) and an outer loop (model fine-tuning). In the inner loop, workflows defined by prompt text (roles, instructions, contracts, hops) are optimized like an evolutionary algorithm. In the outer loop, the top 15 trajectories from tasks performed with the optimized workflows are selected to fine-tune the model using LoRA. During training, the orchestrator's initial instructions are removed, retaining only sub-agent instructions to focus on the task itself.

Experimental Results and Performance

Experiments based on Qwen3.6-27B showed performance improvements on research benchmarks. Score multipliers per output token increased by 1.60x on MLR-Bench, 1.35x on DSBench, and 1.20x on ScienceAgentBench. Conversely, scores dropped to 0.96x and 0.94x on software benchmarks such as Terminal-Bench 2.0 and SWE-bench Verified, revealing limitations in transferability. In synthetic task tests, the second-cycle (L2) model demonstrated a 69.9% win rate compared to the baseline model (L0).

Analysis and Limitations

Multi-agent trajectory learning proved token efficiency by reducing supervised tokens by 31.4% compared to single-agent learning while increasing the win rate by 4.3 percentage points. Additionally, a 'role generalization' phenomenon was confirmed, where learning only task-solving trajectories also improved workflow optimization and evaluation capabilities. However, limitations were noted, including low statistical significance of improvements in some benchmarks, performance drops on software tasks, reliance on external judge models, and safety concerns regarding shared blind spots in homogeneous loops.