Yoonho Lee, Roshen Nair, Qizheng Zhang, and 3 others
Introduces an outer-loop coding agent that optimizes executable LLM harnesses by inspecting the full code, scores, and traces of prior candidates; one evaluation uses LawBench criminal-charge classification.
Limitation Legal evidence is limited to one LawBench criminal-charge classification benchmark within a broader three-domain AI-systems evaluation, and LegalArxiv has not independently reproduced any reported result.