News · 2026-10-12

Sakana’s MASS improves a bounded agent loop, with uneven coding transfer

Sakana AI’s MASS research couples multi-agent workflow search with training of the model that executes, optimizes and judges those workflows. The paper reports improved performance on held-out research tasks after two cycles, while coding efficiency tests remain flat to slightly worse. It is a concrete bounded self-improvement experiment whose uneven transfer limits broader claims.

Key facts

• The reported win share against the bare-prompt base rose from 53.9% after one cycle to 69.9% after two on three synthetic test tasks.
• The paper was submitted October 8 and is titled “Recursive Self-Improvement through Multi-Agent Self-Supervision.”
• Hyunin Lee and colleagues describe a shared-model system at Sakana AI.
• Primary evidence comes from the paper, project explainer and code.

Self-improvement becomes a more useful claim when the object being improved is named. MASS changes two objects: the instructions and structure of an agent team, and the underlying model’s learned behavior. That distinguishes it from systems that only write notes, revise a prompt or add more model calls without any weight update.

The inner loop searches workflows. A workflow defines agent roles, instructions, output contracts and call order. Copies of a shared model execute it, evaluate the resulting work against the current champion and propose revisions. The selected workflow is then run repeatedly to generate trajectories: records of what the orchestrator and individual agents did.

The outer loop trains the shared model on selected team trajectories. The researchers remove explicit workflow text from the orchestrator’s training prompt so the student is intended to internalize coordination rather than merely reproduce a supplied plan. The updated weights then return as solver, optimizer and evaluator in the next cycle. The full paper documents these coupled steps.

A concrete analogy is a project team that first rearranges its working process, then trains its members on the most successful examples of that process. Better organization and learned habits can both improve the next project. But if the same team also judges whether its work is good, shared misunderstandings can survive the whole cycle. MASS makes that risk technically relevant because one model fills several roles.

The synthetic suite contains 12 research-programming tasks spanning finance, robotics and pharmacy. Nine were designated for training and three for testing. One training task failed to yield an outperforming workflow and was excluded from teacher-data collection, leaving eight tasks for post-training. These distinctions matter: the study is not a large, independently sampled test of arbitrary scientific projects.

On the three synthetic test tasks, the reported win share rises from 53.9% after one cycle to 69.9% after two against the bare-prompt base. Those are externally model-judged task-output comparisons reported by the authors. They are not independent human assessments of scientific correctness or economic value. Four public research benchmarks show improved score per output token, according to the paper.

The coding results are essential counterevidence. On Terminal-Bench 2.0 and SWE-bench Verified, the second-cycle score-per-output-token ratios are 0.96 and 0.94 of the base. The same update does not produce a universal improvement across the tested task families. This is a useful distinction from the circulating headline that recursive self-improvement has simply arrived.

Efficiency also needs the right denominator. A separate control reports that the multi-agent student used about 31.4% fewer processed supervised tokens than the single-agent control while reaching a higher win share. That calculation excludes the full cost of workflow search, teacher generation and evaluation. It is a supervised-data efficiency result, rather than a verified reduction in total compute or dollars.

The study starts from a Qwen3.6-27B model. The dossier does not verify a downloadable checkpoint file total or a primary-source running-memory requirement; the VRAM requirement is unstated in the available evidence. Parameter count is not a safe substitute for either measurement. The public code is research code and does not by itself establish a packaged, independently validated product.

The researchers’ ablation also suggests that weak workflow optimization can bottleneck the loop more severely than weak evaluation alone. Search can regress, and the paper identifies shared blind spots as a safety weakness. The existing lessons on agent harnesses and recursive self-improvement give context for those distinct mechanisms.

The caveat is therefore central to the result: MASS provides encouraging evidence for a designed improvement loop on bounded tasks, with model-based judgments and no independent replication established in the dossier. Its significance is the explicit connection between workflow search and weight learning. The next question is whether that connection produces durable gains on unfamiliar tasks under evaluations the loop did not help construct.

Primary source, verified: read the paper → (arXiv 2610.12176)

Ground Truth. (2026, October 12). Sakana’s MASS improves a bounded agent loop, with uneven coding transfer. Ground Truth. https://groundtruth.day/news/sakana-mass-workflows-weights-uneven-transfer.html

@misc{groundtruth:sakana-mass-workflows-weights-uneven-transfer,
  title  = {Sakana’s MASS improves a bounded agent loop, with uneven coding transfer},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/sakana-mass-workflows-weights-uneven-transfer.html}
}

Topics: agents · multi-agent · self-improvement · post-training · evaluation