Harvey post-trains RLM agents for M&A diligence, lifting benchmark pass rate from 23% to 62%
lateinteraction · x · 2026-09-09
Legal AI company Harvey, partnering with Baseten, published research on post-training recursive language model (RLM) agents to run end-to-end M&A diligence.
Key points:
- They built an RLM harness where a root agent searches the data room, delegates document review to sub-agents, and orchestrates a final diligence memo
- Distributed review lets it work through data rooms of up to 80M tokens
- On their synthetic LAB Diligence benchmark, the harness raised average rubric pass rates from 23% to 62%
- A key finding: model-harness co-optimization matters — the root agent processes only 1-4% of tokens (sub-agents handle 96-99%) yet its choice still materially moves performance
Strong evidence that co-optimizing model and harness pays off in long-horizon real-world tasks.
Related event: Harvey's post-trained RLM agent boosts M&A due diligence pass rate(3 posts)→
More from coding & agent
- Are agent swarms more legible than single agents? Safety researchers debate — jd_pressman · 2026-09-09
- Idea: AI coding tools should pitch app ideas from your unused credits and build them — thisiskp_ · 2026-09-09
- AI Can Solve Navier-Stokes but Can't Write a Maintainable TypeScript Abstraction — burny_tech · 2026-09-09
- Open-source MCP server compresses agent context: 93.8% recall vs 51.9% baseline — Beginning_Note_5385 · 2026-09-09
- How one engineering org tripled output in 18 months by removing handoffs, not just AI — rseroter · 2026-09-09
- 14B open model on one RTX 4090 matches hosted frontier model on text-to-SQL; mnemiq open-sourced — ycombinator · 2026-09-09