UCSB and Microsoft Researchers Introduce ScholarEvolve, a Framework That Turns Research Papers Into Agent Harness Upgrades

Researchers from UC Santa Barbara and Microsoft have introduced ScholarEvolve, a framework that automatically mines new information from research papers and turns it into code upgrades for an agent's harness — the component stack that manages prompts, context, skills, memory and workflows — without retraining the underlying model.

In benchmark tests, ScholarEvolve improved the open-weight Qwen3.5-27B from 69.0% to 81.4% task completion on the AppWorld Normal test set, closing roughly 74% of the performance gap with GPT-5.4 without modifying the model's weights. The framework splits the harness into five modules with strict interfaces: tool interface, context management, skills, memories, and agentic workflows. Its pipeline runs four stages — failure auditing and literature retrieval, topic modeling, blueprint and health-probe generation, and module crossover with validation-based selection.

The approach also lifted performance on other tasks. On the AppWorld Challenge, the evolved Qwen3.5-27B harness improved zero-shot generalization from 49.6% to 63.6%, and on tau-Bench Telecom, GPT-5.4-mini rose from 72.7% to 81.9%. The authors note that safety vetting of retrieved papers and third-party code is needed before production use, and that harness evolution complements rather than replaces deterministic policy guardrails.