The Critic Can Hurt: On the Safety of Critics in LLM-Guided AutoML Search
Organizations: Amazon · Independent Researcher · University of Massachusetts Boston · University of Southern California
Abstract
LLM critics increasingly steer iterative search in agentic AutoML, prompt optimization, and self-refinement--evaluated for whether they help, but rarely for whether they can hurt. We isolate the critic's marginal effect through a four-arm audit and characterize the loop's tolerance through controlled fault injection. We show a naive 8B critic drives search below the no-critic random-search floor on 7 of 8 tabular datasets, losing up to 0.245 AUC in the worst case; the magnitude varies strongly with dataset, frontier, and budget. When: stronger critics attenuate the loss on the evaluated suite, without establishing a universal capability threshold. Why: controlled fault injection shows that non-directional errors are nearly costless under best-so-far search, whereas directional errors consume finite search opportunities; an idealized effective-sample calculation yields a necessary condition, , not a deployment certificate. How: a simple observable validation gate and a learned detector can reduce realized loss, but neither is guaranteed to preserve the full-budget no-critic baseline. We will release CRITFAULT, to our knowledge the first benchmark for detecting harmful critic feedback in AutoML. A critic is not a free addition: audit before trusting, and gate when capability or verification is insufficient.
Figures & tables
| 3B 25 | 8B 24 | 8B 25 | 14B | 32B | 70B | H4.5 | |
|---|---|---|---|---|---|---|---|
| adult | |||||||
| electr. | |||||||
| jm1 | |||||||
| kc1 | |||||||
| bank-mkt | |||||||
| credit-g |
| Setting (space size) | 95% CI | Verdict | |
|---|---|---|---|
| Tabular AutoML (96) | harm | ||
| Credit scoring (96) | neutral | ||
| 20NG text-pipe (36) | neutral | ||
| 20NG weak strong (50) | neutral | ||
| M4 forecasting (13) | helps | ||
| RAG retrieval (54) | helps |
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Analysis | Datasets | Purpose |
|---|---|---|
| Real-critic audit | 8 binary | Below-floor harm |
| Fault injection | 9 (see text) | Error structure |
| Capability ladder | 6 targeted | Capability attenuation |
| CritFault | 12 total | Feedback detection |
| Nine datasets (mean , all ) | |||
| bank-marketing | 0.990 | letter | 0.983 |
| credit-g | 0.989 | churn | 0.980 |
| phoneme | 0.988 | jm1 | 0.974 |
| adult | 0.987 | kc1 | 0.966 |
| electricity | 0.985 | ||
| Near-zero headroom (mean ) | |||
| Dataset | 3B 25 | 8B 24 | 8B 25 | 14B | 32B | 70B | H4.5 |
|---|---|---|---|---|---|---|---|
| adult | |||||||
| electricity | |||||||
| jm1 | |||||||
| kc1 | |||||||
| bank-marketing | |||||||
| credit-g |
| Dataset | no-critic | oracle | AutoGluon |
|---|---|---|---|
| electricity | .9411 | .9542 | .9316 |
| credit-g | .8017 | .8085 | .7849 |
| churn | .9412 | .9463 | .9264 |
| adult | .9287 | .9302 | .9236 |
| bank-marketing | .9297 | .9328 | .9309 |
| kc1 | .8279 | .8314 | .8378 |
| Dataset | headroom | @96 | @250 |
|---|---|---|---|
| electricity | [ ] | [ ] | |
| credit-g | [ ] | [ ] | |
| kc1 | [ ] | [ ] | |
| jm1 | [ ] | [ ] | |
| adult | [ ] | [ ] |
| Data | raw nc | filter nc | KL | KL |
|---|---|---|---|---|
| adult | ||||
| jm1 | ||||
| kc1 | ||||
| credit-g |
| Search policy | mean harmful | /12 |
|---|---|---|
| sequential | 9 | |
| multi-agent (pooled) | 9 | |
| tree-search (UCB1) | 9 | |
| parallel -wide | 1 |
| Family (#calls) | adult | jm1 | credit-g | kc1 |
|---|---|---|---|---|
| no-critic (0) | ||||
| naive (8) | ||||
| CoT (8) | ||||
| self-consistency (40) | ||||
| + legacy filter (8) |
| Held-out dataset | raw (§ 4 ) | Fine-tuned (LOO) | VGC |
|---|---|---|---|
| adult | |||
| electricity | |||
| jm1 | |||
| kc1 | |||
| churn | |||
| bank-marketing |
| Dataset | Type | ||
|---|---|---|---|
| credit-g | 700 | 300 | binary |
| kc1 | 1476 | 633 | binary (defect) |
| mfeat-factors | 1400 | 600 | multiclass bin |
| segment | 1617 | 693 | multiclass bin |
| splice | 2233 | 957 | 3-class bin |
| churn | 3500 | 1500 | binary |
| Detector | LODO mean AUC |
|---|---|
| random | |
| majority (base rate) | |
| best single feature | |
| logistic (5 feat.) | |
| gradient-boosted (ours) |
| Dataset | cost | cost | |
|---|---|---|---|
| credit-g | 0.6 | 0.093 | 0.005 |
| adult | 0.6 | 0.084 | 0.001 |
| churn | 0.6 | 0.078 | 0.004 |
| jm1 | 0.9 | 0.061 | 0.004 |
| electricity | 0.9 | 0.051 | 0.002 |
| kc1 | 0.9 | 0.048 | 0.004 |
| Dataset | no-crit | oracle | No-Critic [95% CI] |
|---|---|---|---|
| electricity | .9465 | .9542 | [ ] |
| churn | .9412 | .9463 | [ ] |
| credit-g | .8031 | .8085 | [ ] |
| jm1 | .7281 | .7321 | [ ] |
| bank-marketing | .9298 | .9328 | [ ] |
| phoneme | .9444 | .9483 | [ ] |
| Round | Proposed family | Val AUC | Below incumbent? |
|---|---|---|---|
| 1 | lightgbm | 0.930 | — (initial) |
| 2 | lightgbm | 0.927 | yes |
| 3 | lightgbm | 0.926 | yes |
| 4 | gradient_boosting | 0.922 | yes |
| 5 | hist_gbdt | 0.928 | yes |
| 6 | extra_trees | 0.902 | yes |
Automated Machine Learning
Automated machine learning (AutoML) builds and tunes data-to-model pipelines with minimal manual effort, balancing predictive quality, compute, and deployment constraints.
See all 105 papers