Recursa Proposes Recursively Self-Improving Foundation Models
A paper submitted to arXiv presents Recursa, a foundation model system designed to adapt to unfamiliar domains by learning from its own experience without human experts or a stronger teacher model. Its knowledge repository stores domain-specific skills, instructions for applying them, memories of past tasks, and reflections. Parallel learners execute tasks and propose local improvements, which are validated and merged into the repository. The authors report that on Harvey LAB’s Firm Knowledge benchmark, the held-out test all-pass rate increased from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash.
The paper describes a concrete system for adapting foundation models to new domains through self-generated experience, a structured knowledge repository, and a recursive process of validating and merging improvements.
Computer Science > Artificial Intelligence
arXiv:2611.04035 (cs)
[Submitted on 2 Nov 2026]
Title:Recursa: Toward Recursively Self-Improving Foundation Model Systems
Authors:Erkang Zhu, Yin Lin, Xiaorong Shi, Xuanrui Lin, Jingren Zhou
View a PDF of the paper titled Recursa: Toward Recursively Self-Improving Foundation Model Systems, by Erkang Zhu and 4 other authors
#Recursa#DeepSeek-V4.1-Flash#Qwen3.8-Flash#Harvey LAB Firm Knowledge benchmark