MapReduce-based specialized agent hits 50% all-pass rate on Harvey's Legal Agent Bench
j_foerst · x · 2026-11-19
Quoting grubelai's first research result: through task-specific specialization, their simple agent design based on MapReduce achieves a state-of-the-art all-pass rate of 50% on Harvey's Legal Agent Bench. The approach elevates models on LAB and even beats LAB-specific post-training, showing the power of specialization.
More from coding & agent
- Microsoft's Copilot Fusion tops Terminal Bench, 5 points over Claude at less than half the cost — 0xkarasy · 2026-11-19
- GitHub's Codename MDASH ships 99 verified vuln fixes for OSS used 2.5B+ times a month — 0xkarasy · 2026-11-19
- At MLOpsWorld, only 3-5 people in the room still read code—the 'post-literate engineer' era — annetgriffin · 2026-11-19
- Open-source Switch Buddy puts Claude Code and Codex in one Slack room to code, review and debug together — ThePeterMick · 2026-11-19
- CodeRabbit launches Verify to catch broken features before PRs merge — EXM7777 · 2026-11-19
- Fuzzy UI: letting users answer roughly while agents handle the precision — round · 2026-11-19