Recursa Lets Foundation Models Recursively Improve in New Domains

Introduction

When a foundation model enters an unfamiliar domain, general language ability is only part of the problem. The model may lack domain-specific knowledge, procedural conventions, and a dependable way to identify mistakes. In many real settings, there is also no large collection of expert demonstrations and no stronger model available to provide continuous supervision. A paper from Alibaba researchers introduces Recursa, a system designed to address this gap by learning from its own experience.

How the system works

Rather than treating adaptation as a simple continuation of parameter training, Recursa maintains an evolving knowledge repository. The repository records several kinds of information:

• Domain skills, which describe knowledge or procedures useful for solving particular tasks.
• Application instructions, which indicate when a skill should be used.
• Task memories, preserving experience from earlier attempts.
• Reflections, capturing what went wrong and what should be improved next.

Recursa first bootstraps the repository from materials in the target domain. It also practices on tasks generated from those materials, helping the system address gaps in its domain knowledge. The next stage focuses on what the paper describes as a behavior gap: multiple parallel learners execute tasks, inspect their experience, and propose local improvements. Candidate changes are then validated before being merged into the shared repository.

This design is more than an answer cache. A useful entry includes not only what to do, but also when to do it and how previous attempts should influence future decisions. The repository therefore becomes a working memory for the system’s evolving procedures. As new tasks are completed, the accumulated skills and reflections can shape later execution, while the learning process itself can also be refined. This feedback loop is the basis of Recursa’s recursive self-improvement approach.

Results and implications

On the Harvey LAB Firm Knowledge benchmark, Recursa increased the held-out all-pass rate for DeepSeek-V4.1-Flash from 30.0% to 55.0%. With Qwen3.8-Flash, the rate rose from 35.0% to 46.0%. The reported results suggest that structured memory, parallel experimentation, and validation can improve domain adaptation even without direct supervision from human experts or a stronger teacher model.

The broader significance lies in the separation between the foundation model and an external, evolving knowledge layer. The model performs reasoning, execution, and reflection, while the repository supports long-term accumulation and reuse. This could be useful for agent systems that repeatedly operate in a specialized environment, where a one-off prompt or static context is not enough.

The evidence should nevertheless be interpreted carefully. The supplied material reports results on one benchmark and does not establish that recursive improvement will be equally reliable across domains. Error accumulation, the quality of self-generated feedback, the validation of proposed changes, and repository growth remain important open questions. Recursa is therefore best viewed as a system design direction for continual adaptation, rather than proof that models can improve safely without external oversight in every setting.

Source: Hugging Face Daily Papers