Research · Updated 12 Oct, 11:28 pm IST
Sakana AI and UC Berkeley propose MASS method for recursive self-improvement
Why it matters for readers: Curious readers should know MASS experiments show promise for self-improving AIs while also revealing fundamental limits when a model judges itself.
- MASS uses copies of a single language model to divide roles: solving tasks, judging results, and refining workflows.1
- Workflows are encoded as text appended to prompts; a contract specifies required outputs and checks, and a hop orders calls between sub-agents.1
- An orchestrator handles distributing work between the sub-agents created from the same model.1
- The researchers observe that using the same model to judge its own outputs can miss mistakes, so shared blind spots may be reinforced across training cycles.1
- The paper notes that some verification tasks at the research frontier can be extremely costly, citing single studies that required more than 120 million CPU hours.1
Get a brief like this every morning
Uzha reads hundreds of sources and gives you the stories that matter for your work, with every source linked. Free.
Get started