Post-Training RLM Agents for End-to-End M&A Diligence (www.harvey.ai)

Recent research led by Baseten has introduced a recursive language model (RLM) harness that significantly improves AI capabilities in end-to-end M&A diligence tasks, which typically involve massive datasets containing up to 80 million tokens. This innovative approach allows for a root agent to delegate tasks to sub-agents that process segments of the data room independently, resulting in a remarkable 39.1 percentage point increase in rubric pass rates across seven evaluated models. The implementation shows that post-training, particularly using reinforcement learning, can further enhance performance; for instance, a Qwen3.5 orchestrator achieved a pass rate improvement from 29.9% to 63% after post-training, indicating substantial advancements in AI handling complex legal tasks efficiently. The significance of this research lies in its potential to transform how AI tools can be adapted for document-intensive legal workloads, like M&A transactions, where thoroughness and context understanding are critical. With improved coverage of data rooms and optimized division of labor among agents, this approach not only increases the quantity of examined documents but also enhances the quality of output, as shown by higher adherence to detailed rubric criteria. Future experiments aim to scale this RLM harness and explore its applicability in other legal contexts, providing a promising avenue for enhancing AI's functionality in high-stakes environments.