MASS

What it is

MASS is research on how a model can improve itself without a stronger teacher. The researchers argue that in research areas that go beyond accumulated human knowledge, even experts may struggle to evaluate results, so they start from a setting in which the model being improved is itself the most capable supervisor. They call this homogeneous recursive self-improvement (homogeneous RSI). The core idea of MASS is to use a team's collective intelligence, which a single model cannot reliably produce on its own, as the training signal. The researchers released both the paper and the code.

How it works

The same language model plays three roles. The task executor produces code, results and reports; the workflow optimizer reads the evaluation feedback and revises how work is divided and how results are passed along; and the evaluator compares two outputs and picks the better one. In the inner loop, the model weights stay fixed while the workflow (roles, instructions, a "contract" defining what information is exchanged, and call order) is varied, and whenever a candidate beats the current best "champion" workflow, it replaces it. In the outer loop, the champion workflow is run many times, the model ranks the results to pick out good execution trajectories, and these are distilled into the model with LoRA. The updated model then takes on the three roles again and the process repeats.

Why it matters

What stands out is that improvement happened through self-evaluation alone, without an external grader, even on research-style tasks whose answers cannot be checked automatically. The researchers report that after applying it twice to Qwen3.6-27B, the win rate against the base model on held-out synthetic research tasks rose from 53.9% to 69.9%, and the score per output token on four research benchmarks increased by 1.2–1.6x. They also report that although the model was trained only on task-execution records, its workflow optimization and evaluation abilities improved as well. They further report that records of multiple agents working together achieved better results with less training data than single-agent records.

Things to know

The figures are results reported by the researchers themselves, and the technique is still at the research stage. The improvements did not carry over consistently to software engineering benchmarks. Because the same model both produces and evaluates the work, there is a risk that blind spots shared by the model get repeatedly reinforced in subsequent training. The researchers also note that supervision capable of correcting self-evaluation is needed.

Example

In a finance task the researchers released, the workflow refined by MASS divided work among sub-agents in the order "data preparation → feature construction → modeling → backtesting → audit → cross-validation → report." The "contract" specified that the data agent pass on to the next agent the results of checking that no future information had leaked into past decisions. By retraining the model on such well-organized team work records, the aim is for the model to eventually learn to organize the work just from the task, without being given a workflow.

Related terms

Multi-agent Agentic workflow LoRA Distillation LLM-as-a-Judge Post-training

References

• Recursive self-improvement through collective intelligence (GeekNews)
• An exact rotating black hole solution in Palatini Kalb-Ramond gravity (arxiv.org)
• Sakana AI Introduces MASS for AI Self-Improvement (Techiexpert.com)
• MASS: Recursive Self-Improvement Through Multi-Agent Self-Supervision (pub.sakana.ai)