UChicago Paper Finds LLMs Cannot Anticipate Their Own Cruxes

A University of Chicago paper shows that LLMs cannot reliably predict which sub-questions actually drive their differing answers; such cruxes can only be identified by comparing multiple sampled predictions, though the study tested only four weaker models.

2026-10-22 ~ 2026-10-24 · 4 related posts

• New arXiv paper extracts the cruxes of LLM forecasts by contrasting 30 predictions — universeinanegg · 2026-10-22
• UChicago paper: LLMs can't anticipate their own cruxes — sampling many rollouts is unavoidable — rohanpaul_ai · 2026-10-24
• arXiv: 'Can LLMs Anticipate Their Cruxes?' — a priori rationales fall short, rollouts required — rohanpaul_ai · 2026-10-24
• U Chicago Paper: LLMs Can't Predict Their Own Answer Split Points, Only 4 Weak Models Tested — askerlee · 2026-10-24