← Back to live feed1 story
Sakana AI and UC Berkeley developed the Multi Agent Self Supervision (MASS) method, which uses virtual subagents to train a 27B model, increasing scores per output token by 1.2-1.6x across four research benchmarks. The approach addresses a supervision bottleneck in complex tasks but raises safety concerns about models reinforcing their own blind spots.
Sign in to suggest edits
Key sources
- SOURCE@sakanaailabs“score per output token reached 1.2-1.6x the base model's level across four research benchmarks”x.com
- SUPPORT@hardmaru“After two cycles with a 27B open-weights model, performance improved 1.2-1.6x per output token on tough research benchmarks (MLR-Bench, ScienceAgentBench, DSBench)”x.com
- SUPPORT@nptacek“been sayin'”x.com
- SOURCEmarketbrief.now