Sakana AI has announced MASS, a research method in which copies of one AI model divide up a task, assess their work and turn selected attempts into training material. The collaboration with UC Berkeley tests a difficult question: can a model improve when it also supplies the feedback? The company’s page is dated 12 October; the preprint appeared on 8 October.
Based on 6 reviewed sources · See sources
The team’s work becomes the next lesson
A finance task illustrates the idea. One agent prepares data, another builds a model and others test and audit it. They share the same underlying Qwen3.6-27B model, but receive different assignments and instructions about what to pass to the next agent.
According to the paper, MASS first searches for better ways to organise that work while the model stays unchanged. It then trains the shared model on selected records of the team’s actions. The updated model returns as worker, organiser and evaluator. The researchers tested two complete cycles—not an indefinitely running learning system. Sakana’s announcement below links the paper and code.
Sakana introduces MASS and its collaboration with Berkeley. The reported results come from the research authors; this is a research announcement.
View the original post on X ↗More efficient research, uneven transfer
The authors report higher benchmark scores per output token—the chunks of text the model generates—on four research tests after two cycles. That measure combines the task score and the amount of generated text. It is not a measure of the full training bill or a percentage of answers correct.
The detailed results show that the improvement did not carry over evenly: the same measure fell slightly on two software tests. Completion also matters. On AstaBench, missing reports receive zero, and more reports were delivered after training. A better aggregate score therefore does not establish that every completed report became better.
The gain depends on the task
Authors’ reported score per output token after two MASS cycles. Each benchmark’s base model = 100%.
Who checks the checker?
The experiment separates learning from measurement. Qwen’s own judgments guide the learning loop; stronger external models, GPT-5.5 and Claude Opus 4.8, assess the synthetic research outputs separately. Those assessments are withheld from MASS. An outside judge used by the researchers is still part of their evaluation, rather than an independent replication.
Aron Brand, CTO of enterprise-data company CTERA, raised the shared-blind-spot concern in the earlier preprint discussion. His comment is a practical judgment, not a reproduced MASS test:
“With the same base model on both sides errors compound silently, so I'd put a different model in the grader seat.”
Sakana itself warns that a shared error could be accepted as progress and reinforced through training. Its public code provides a route to investigate the method; running real experiments requires model serving and GPU setup. The useful next evidence would show whether the gains survive further cycles and independent checks of the work produced.
Sources
- Recursive Self-Improvement through Collective Intelligence
Sakana AI · Announcement ·
- MASS method and reported results
Sakana AI and UC Berkeley researchers · Research
- MASS preprint and evaluation protocols
Hyunin Lee and co-authors on arXiv · Research ·
- MASS research implementation
Sakana AI on GitHub · Documentation
- Original research announcement
Sakana AI on X · Announcement ·
- Shared-grader concern in the preprint discussion
Aron Brand on X · Research ·
Written with AI assistance from the sources above. Analysis and practical implications are our interpretation. Our editorial approach.
