UCSB and Microsoft release ScholarEvolve self-improving agent framework
ScholarEvolve has agents read arXiv papers and check official code, then rewrite their own configuration; Qwen3.5-27B rose to 81.4% on AppWorld.
Researchers at the University of California, Santa Barbara and Microsoft have released a framework called ScholarEvolve that lets AI agents read academic papers and modify their own code to improve task performance, according to a paper posted to the arXiv online archive.
ScholarEvolve searches for papers on arXiv, has a research agent read them and check the official code that accompanies them, and then hands the findings to a coding agent that implements changes. The framework splits an agent into five modules: tool interface, context management, skills, memory and workflow. Candidate changes are evaluated on validation data and only replace the existing configuration when the improvement is confirmed across the whole system.
Benchmark results
On the AppWorld normal test, Qwen3.5-27B's task completion rate rose from 69.0% to 81.4%. The researchers say this cut the performance gap with GPT-5.4 by about 74%. On the tau-Bench Telecom test, GPT-5.4 mini improved from 72.7% to 81.9%.
On the AppWorld Challenge test, which uses unseen Amazon and Gmail APIs, the Qwen3.5-27B agent rose from 49.6% to 63.6%, scoring 9 percentage points higher than Meta-Harness on the same test. In cases involving Python execution errors, the improved agent completed 59.9% of tasks against 44.9% for the original configuration.
The team ran three rounds of evolution while keeping the Qwen model weights and the existing task data unchanged, and AppWorld performance improved in each round.
Quick answers
What is ScholarEvolve?
A framework from UC Santa Barbara and Microsoft researchers that lets an AI agent search arXiv, read papers, check official code and implement changes to its own configuration.
How much did performance improve?
On the AppWorld normal test, Qwen3.5-27B went from 69.0% to 81.4%; on tau-Bench Telecom, GPT-5.4 mini went from 72.7% to 81.9%.
Were the model weights retrained?
No. The three rounds of evolution were run while keeping the Qwen model weights and existing task data unchanged.