SakanaAI/mass

What it solves

MASS addresses the challenge of improving AI model performance by creating a self-improving loop. Instead of relying solely on static training data, it allows a model to iteratively refine its own workflows and learn from the successful trajectories it generates.

How it works

The system operates in a recurring cycle. First, it uses a process called RHI to search for and optimize multi-agent workflows while the model's core weights remain fixed. It then collects the successful trajectories from these workflows and uses them to perform post-training (fine-tuning) on the model. This updated model then becomes the starting point for the next cycle, creating a recursive improvement process.

Who it’s for

This project is for researchers and developers interested in recursive self-improvement, multi-agent systems, and automated post-training pipelines. It provides a research-grade implementation for those looking to experiment with iterative model updates using synthetic tasks and benchmark evaluations.

Highlights

• Automated pipeline for workflow search, trajectory collection, and model ranking.
• Integrated post-training tools including LoRA, assistant-only loss masking, and model export.
• Includes 12 synthetic tasks across finance, pharmacy, and robotics for testing.
• Modular command structure allowing users to inspect, pause, and resume specific stages of the improvement cycle.
• Support for evaluating model progress against public benchmarks like ScienceAgentBench and MLR-Bench.