Training Frontier Review Table Models With Applied Compute

This post details the training of a specialized model for Harvey's Review Table product in partnership with Applied Compute. Key contributions include: the development of a synthetic dataset approximating real-world legal document analysis tasks, a novel reward function design that balances answer correctness, schema adherence, and citation quality (precision and recall), and the empirical results showing the custom-trained model's superior performance (0.903 Answer Score, 84.1% citation precision, 91.93% citation recall) and cost reduction (54.8% lower cost per cell) compared to frontier models. It also introduces an experiment comparing single-turn semantic retrieval with a multi-turn agentic retrieval harness for evidence selection.