Sakana AI Unveils MASS: Recursive Self-Improvement via Multi-Agent Self-Supervision, Boosting Win Rates to 69.9%

Background and the Problem

Recursive Self-Improvement (RSI) is fundamental to autonomous intelligence. While reinforcement learning with verifiable rewards (RLVR) flourishes on formal programming and mathematics, open-ended scientific discovery and non-verifiable reasoning tasks lack programmatic oracles. Furthermore, single-agent models evaluating their own outputs succumb to homogeneous confirmation bias, unable to recognize their own blind spots without external supervision.

Architecture and How It Works

To break the supervision bottleneck in non-verifiable domains, Sakana AI introduces MASS (arXiv:2610.12176, site: sakana.ai/mass):

1. Single Base Model with Multi-Agent Specialization: Instantiated on a single foundation model (e.g., Qwen3.6-27B), MASS tasks the network with three complementary duties: execution workers, an evolutionary workflow optimizer that discovers novel agentic topologies, and blind evaluators.
2. Alternating Evolution and SFT Cycles: The self-improvement loop alternates between exploring multi-agent interaction spaces to optimize collaboration graphs, and executing supervised fine-tuning (SFT) on the best self-generated execution trajectories, feeding upgraded weights back into the next evolutionary round.
3. Diversity Amplification: Multi-agent debate and cross-examination break single-model echo chambers, synthesizing dense, exploratory reasoning data that prevents policy collapse.

Benchmarks and Measured Results

Evaluated across complex scientific and reasoning suites (MLR-Bench, ScienceAgentBench, DSBench, AstaBench):

1. Win Rate Surges to 69.9%: Against the unadapted base model, win rate increases from 53.9% after cycle 1 to 69.9% after cycle 2. Second-round checkpoints decisively defeat first-round checkpoints in 60.8% of head-to-head trials.
2. 1.2x to 1.6x Token Efficiency: Output efficiency per generated token rises across all research benchmarks, demonstrating crisper deductive reasoning without redundant verbosity.
3. Outperforms Single-Agent Training with 1.4x Less Data: A student model trained on multi-agent MASS traces surpasses a single-agent student trained on 1.4x larger token budgets.

Getting Started for Developers

The MASS research and evolutionary topologies are detailed on Sakana AI's portal (sakana.ai/mass). Developers and AI researchers building autonomous research agents can implement alternating workflow evolution and self-distillation to achieve recursive capability improvements in complex domains lacking rigid compilers or verifiable reward functions.