Recursa recursively self-improving foundation model system
Original title: Recursa: Toward Recursively Self-Improving Foundation Model Systems
Recursa adapts to new domains by learning from its own experience and raises all-pass rates on held-out tasks from 30.0% to 55.0% with DeepSeek-V4.1-Flash.
Recursa uses a knowledge repository storing domain-specific skills, instructions, task memories, and improvement reflections. It bootstraps the repository from domain materials and refines it through recursive self-improvement. Parallel learners execute tasks, synthesize local improvements, and validate merges into the repository.
- Raises all-pass rate on Harvey LAB's Firm Knowledge benchmark from 30.0% to 55.0% with DeepSeek-V4.1-Flash
- Raises rate from 35.0% to 46.0% with Qwen3.8-Flash
- No human experts or stronger teacher model required
- Bootstraps from domain materials and refines via recursive self-improvement
- Organizations
- Harvey LAB
- AI models
- DeepSeek-V4.1-FlashQwen3.8-Flash
From the original article
1 min read at huggingface.co
Foundation models in changing environments can encounter unfamiliar domain knowledge that their original training did not cover. However, new domains rarely provide large amounts of expert-created training data or reliable supervisory signals for further training.
By Erkang Zhu, Yin Lin, Xiaorong Shi · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.