From Sakana AI · · 1m 50s
Multi-Agent Self-Supervision for AI Self-Improvement
AI system learns to improve itself using team coordination.
In brief
AI system learns to improve itself using team coordination. AI model enhances research capabilities through self-supervised multi-agent teamwork. Originally reported by Sakana AI.
MASS: Recursive Self-Improvement
AI system improves itself using multi-agent teamwork to create better research outputs.
Why Self-Supervision Matters
Frontier AI models need self-supervision to advance beyond human knowledge without external evaluators.
Homogeneous Recursive Self-Improvement
The model acts as both optimizer and evaluator to improve itself through self-supervision.
Teamwork for AI Learning
MASS uses multi-agent teams to create self-supervision signals for recursive self-improvement.
How MASS Works
Inner loop optimizes team workflow while outer loop updates model weights with selected executions.
Workflow Optimization
Team workflow search finds optimal sequences for data preparation, feature construction, and backtesting.
Training Process
MASS trains on selected team executions to improve model performance across research tasks.
Performance Results
Two rounds of MASS improve task performance from 53.9% to 69.9% against base model on test tasks.
Benchmark Improvements
MASS achieves 1.2, 1.6× better performance on public benchmarks after two rounds of self-improvement.
Role Generalization
MASS improves workflow optimizer and evaluator capabilities through self-supervision.
Coordination Without Workflow
Model learns to delegate tasks without explicit workflow instructions through self-improvement.
Information Flow Analysis
Optimal workflows encode more task information in contracts and call order than roles and instructions.
Trajectory Effectiveness
Multi-agent trajectories provide denser learning signals than single-agent trajectories for self-improvement.
Optimizer Bottleneck
The model's optimizer is the main bottleneck for self-improvement, not the evaluator.
Safety Concerns
Shared blind spots in self-evaluation could lead to unsafe improvements if not corrected.
The Odyssean Moment
AI could achieve continuous self-learning without external supervisors.
At a glance
AI model enhances research capabilities through self-supervised multi-agent teamwork.
