Amazon paper: weak LLM critics can sink search loops, ideal gain only 0.005 AUC
rohanpaul_ai · x · 2026-11-18
LLM critics may hurt more than help
A new paper from Amazon and UMass Amherst finds that adding an LLM critic to an AutoML search loop—one that reads past results and suggests what to try next—has tiny upside and much larger downside.
Setup
- Tabular model search over a fixed pool of 96 pipeline configurations, run with and without a critic.
- Random critic mistakes cost almost nothing since the loop keeps its best result, but advice that steadily points the wrong way wastes tries.
Key results
- A Llama-3.1-8B critic scored below the no-critic baseline on 7 of 8 datasets, losing up to 0.245 AUC.
- Even an ideal critic gained at most 0.005 AUC.
- The harm mostly faded with 250 setups and 20 rounds.
Takeaway
Before trusting an LLM critic, even a large one, run the same budget without it and drop it if it loses. Teams usually check whether critics help, rarely whether they hurt.
More from coding & agent
- Cognition launches Devin Bots: assign outcomes, not tasks, and agents run for weeks — marvinvonhagen · 2026-11-18
- FastMCP 4.2 ships Skills over MCP — drag a skill folder, get an MCP server instantly — HanchungLee · 2026-11-18
- OpenAI engineers secretly use Astra for Law to write PRDs and design docs — sherwinwu · 2026-11-18
- START-HERE.md: Mapping Your Vault So Agents Find Files Without Burning Tokens — tomcrawshaw01 · 2026-11-18
- AnyMachine turns a single painting into a 160-part walking robot rig built by an agent in Blender — bennash · 2026-11-18
- Agent-run 'company' needed just 6 compactions in 1.5 months; paper shows benchmarks are compaction-friendly — maksym_andr · 2026-11-18