KDD paper: naive 8B LLM critic drags AutoML search below random search, up to 0.245 AUC lost
rohanpaul_ai · x · 2026-11-18
An Amazon + UMass paper accepted at KDD 2027, "The Critic Can Hurt," systematically audits LLM critics in agentic AutoML search loops.
Key findings:
- A four-arm audit isolates the critic's marginal effect: a naive 8B critic pushes search below the no-critic random-search floor on 7 of 8 tabular datasets, losing up to 0.245 AUC, while the upside is only 0.005 AUC at best.
- Controlled fault injection explains why: under best-so-far search, non-directional errors are nearly costless, but directional errors consume finite search opportunities; an idealized effective-sample calculation yields a necessary condition, not a deployment guarantee.
- Stronger critics attenuate losses with no universal capability threshold; a simple observable validation gate and a learned detector reduce realized loss but neither reliably preserves the full-budget baseline.
The authors will release CRITFAULT, the first benchmark for detecting harmful critic feedback in AutoML. Takeaway: critics aren't free — audit before trusting and gate when capability or verification is insufficient.
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
- Amazon paper: weak LLM critics can sink search loops, ideal gain only 0.005 AUC — rohanpaul_ai · 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