arxiv:2611.04035

Recursa: Toward Recursively Self-Improving Foundation Model Systems

Published on Nov 2

· Submitted by

Yin Lin on Nov 4

alibaba

Upvote

Authors:

Yin Lin ,

Abstract

Recursa is a self-improving foundation model system that adapts to new domains by recursively learning from its own experience, storing skills and reflections in a knowledge repository, and validating improvements without human supervision or stronger teachers.

Generated by thinkingmachines/Inkling-Small

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. We introduce Recursa, a foundation model system that adapts to new domains by learning from its own experience without human experts or a stronger teacher model. Recursa uses a knowledge repository to store what it learns: domain-specific skills and instructions on when to apply them, along with memories of past task experiences and reflections on what to improve. Recursa bootstraps the repository from domain materials and then refines it through recursive self-improvement. It uses parallel learners to execute tasks and synthesize local improvements, then validates and merges the improvements into the knowledge repository. On Harvey LAB's Firm Knowledge benchmark, Recursa raises the all-pass rate on held-out test tasks from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash.

View arXiv page View PDF Add to collection

Community

niceirene

Paper author Paper submitter 1 day ago

When a foundation model moves into a new domain, there's usually little expert-written data or reliable supervision to train on. Recursa adapts to the domain by learning from its own experience, without human experts or a stronger teacher model.

A knowledge repository stores what the system learns: domain-specific skills, instructions on when to use them, memories of past tasks, and reflections on what to improve.
Bootstrapping: the system builds the repository from domain materials and practices on tasks it generates from them. This targets the knowledge gap.
Recursive self-improvement: parallel learners run tasks and propose local improvements, which are validated and merged back into the repository. This targets the behavior gap. The system's own model drives the whole process, including improving how it learns.

On Harvey LAB's Firm Knowledge benchmark, Recursa raises the all-pass rate on held-out test tasks from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash.

Paper: https://arxiv.org/abs/2611.04035