RESEARCH · RESEARCH · #1890
Sakana AI and UC Berkeley show single LLM can self-improve via MASS but internal evaluator can reinforce blind spots
Researchers from Sakana AI and UC Berkeley introduce Multi-Agent Self-Supervision (MASS), a method where one language model runs copies of itself as sub-agents to solve, judge, and optimize workflows. Using Qwen3.6-27B as the base, the approach raised win rates versus the base model to ~70% after two rounds on several benchmarks, but shared blind spots in the model’s self-evaluation sometimes accept flawed solutions and propagate errors into subsequent training.
KEY POINTS
- Researchers from Sakana AI and UC Berkeley introduce Multi-Agent Self-Supervision (MASS), a method where one language model runs copies of itself as sub-agents to solve, judge, and optimize workflows.
- Using Qwen3.6-27B as the base, the approach raised win rates versus the base model to ~70% after two rounds on several benchmarks, but shared blind spots in the model’s self-evaluation sometimes accept flawed solutions and propagate errors into subsequent training.
- This shows a scalable path to recursive LLM self-improvement without a stronger teacher but highlights a key safety failure mode: internal evaluators can validate and amplify their own mistakes.
WHY IT MATTERS
This shows a scalable path to recursive LLM self-improvement without a stronger teacher but highlights a key safety failure mode: internal evaluators can validate and amplify their own mistakes.