Assertion 1
Sakana says two rounds of this self-training took a 27 billion parameter model (parameters are a model's learned settings) to a benchmark score, meaning a score on a standard test set, of 1.2 to 1.6 times the base model's level for each token it produced, measured across four research tests, with the training itself done on synthetic open-ended research tasks, which are machine-generated tasks rather than real user work Sakana AI.
Assertion status: No spot-check verdict is published for this assertion.
Sakana AI introduced a method called Multi-Agent Self-Supervision (MASS), where a single shared model uses virtual subagents to solve tasks and selects the best workflows using its own judgments.
No stored spot-check names this claim in this edition.
After two training cycles on synthetic open-ended research tasks using a 27B parameter model, the score per output token reached 1.2 to 1.6 times the base model's level across four research benchmarks.
No stored spot-check names this claim in this edition.
The researchers found that training only on task-solving data from the optimizee naturally improves the capabilities of both the optimizer and the evaluator within the recursive self-improvement loop.
No stored spot-check names this claim in this edition.