Researchers Made AI Agents Better by Turning Academic Papers Into Working Code, No Retraining Required
VentureBeat, Wednesday, October 21st, 2026
ScholarEvolve turns research papers into harness upgrades, lifting agent task completion by up to 12 points without retraining.
UC Santa Barbara and Microsoft researchers introduced ScholarEvolve, which improves an agent's harness by mining academic papers instead of relying only on model knowledge or past traces.
It splits the harness into five modules (tool interface, context management, skills, memories and workflows), audits failed tasks, retrieves relevant arXiv papers, implements candidate techniques and merges the best modules after statistical validation.
Without changing weights, it raised Qwen3.5-27B on AppWorld from 69.0% to 81.4% and GPT-5.4-mini on tau-Bench Telecom from 72.7% to 81.9%.
VentureBeat notes the code repository was empty and advises vetting research-derived changes before deployment.