Recursa: Toward Recursively Self-Improving Foundation Model Systems
| Source: arXiv AI
Tags: Recursa, Recursive Self-Improvement, Harvey LAB, DeepSeek-V4.1-Flash, Qwen3.8-Flash, Knowledge Repository
Recursa, a new foundation model system, adapts to unfamiliar domains by learning from its own experience, lifting held-out all-pass rates on Harvey LAB's Firm Knowledge benchmark from 30% to 55% without human experts or a stronger teacher.
Details
New domains rarely come with large expert-labeled datasets or reliable supervision, which limits how far foundation models can be adapted after training. Recursa stores what it learns in a knowledge repository: domain-specific skills, instructions on when to apply them, memories of past tasks, and reflections on what to improve. It bootstraps the repository from domain materials, then refines it through recursive self-improvement, with parallel learners executing tasks and proposing local improvements that are validated and merged. On Harvey LAB's Firm Knowledge benchmark, the all-pass rate on held-out tasks rose from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash. For teams adapting models to legal or other specialist work, the approach changes what the system knows at inference time rather than retraining weights.