Under target-blind LLM supervision, every learner’s worst-case risk is at least half the model’s admissible-label overlap mass at every sample size, and that floor is certifiable from held-out unlabeled inputs.
Language-models-as-a-service: Overview of a new paradigm and its challenges
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
NL-PAC: Specification Ambiguity and Certified Minimax Risk Floors in LLM-Mediated Supervision
Under target-blind LLM supervision, every learner’s worst-case risk is at least half the model’s admissible-label overlap mass at every sample size, and that floor is certifiable from held-out unlabeled inputs.