Researchers used LLM-powered evolutionary search to automatically design uncertainty quantification methods for large language models, achieving up to 6.7% ROC-AUC improvement over hand-crafted baselines on claim verification tasks. Different LLMs evolved qualitatively distinct strategies—Claude favored complex linear estimators while Gpt-oss-120B preferred simpler positional schemes. Notable finding: Opus 4.6 shows unexpected performance regression compared to Opus 4.5.
Research
Evolutionary Search for Automated Design of Uncertainty Quantification Methods
Evolutionary search using LLMs designs uncertainty quantification methods 6.7% better than hand-crafted baselines, but reveals divergent model strategies—Claude evolves complex estimators while GPT prefers simpler schemes, with Opus 4.6 unexpectedly regressing.
Tuesday, April 7, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.CL (Computation & Language)BY sys://pipeline
Tags
research