Recursa: Toward Recursively Self-Improving Foundation Model Systems

Recursa. Toward Recursively Self-Improving Foundation Model Systems. By Erkang Zhu and four others, from Alibaba Research. Published on the second of November.

A foundation model deployed in a changing environment can run into domain knowledge its original training never covered. The usual remedy is more training, but new domains rarely come with large amounts of expert-created training data, or with reliable supervisory signals to train against. That leaves a system facing unfamiliar material with no one to learn from but itself.

Recursa is a foundation model system that adapts to a new domain by learning from its own experience. It needs neither human experts nor a stronger teacher model. What it learns is stored in a knowledge repository, and that repository is what keeps improving.

The repository holds four kinds of thing: domain-specific skills, instructions on when to apply them, memories of past task experiences, and reflections on what to improve. Recursa first bootstraps it from domain materials. It then refines it through recursive self-improvement. Parallel learners execute tasks and synthesize local improvements, and those improvements are validated before being merged back into the shared repository.

The loop is recursive in a plain sense. The repository shapes how tasks are carried out, the experience of carrying them out produces improvements, and the validated improvements change the repository for the next round. The skills say what to do, the instructions say when, the memories record what happened, and the reflections say what should change.

The test is Harvey LAB's Firm Knowledge benchmark, measured on held-out test tasks by the all-pass rate. With DeepSeek-V4.1-Flash, Recursa raises that rate from thirty percent to fifty-five percent. With Qwen3.8-Flash, it rises from thirty-five percent to forty-six percent. Both gains are measured on tasks held out from the learning, so they reflect what carries over rather than what was practised.

The learning here is written down. Skills, instructions, memories and reflections accumulate in a repository the system writes for itself, checked before anything is merged, and the gains show up with two different underlying models.