U Chicago Paper: LLMs Can't Predict Their Own Answer Split Points, Only 4 Weak Models Tested
askerlee · x · 2026-10-24
A new University of Chicago paper suggests an LLM's own account of what drives its decisions is a poor guide to what actually varies across its answers. LLMs fail to predict the key sub-questions ("cruxes") their answers split on—these only surface when comparing many answers—so simply asking a model what it cares about yields misleading introspection. Askerlee adds a caveat: the study only evaluated gpt-5.4-nano, qwen3:32b, gemma4:31b and gpt-oss:120b, all fairly weak models, so the finding's applicability to stronger models remains uncertain.
Related event: UChicago Paper Finds LLMs Cannot Anticipate Their Own Cruxes(4 posts)→
More from Research
- Explicit Diophantine equation provably equivalent to Con(ZFC) constructed, Lean proof in progress — IgorCarron · 2026-10-24
- Study Estimates LLMs Generate $123-$417 of Economic Value per Megajoule — dan_s_becker · 2026-10-24
- AI-Assisted Paper on Z-Library Cites a TorrentFreak Article That Doesn't Exist — TangerineDream · 2026-10-24
- A new paper offers a framework for measuring firms' AI efforts, datasets included — TaniaBabina · 2026-10-24
- LLM inference may be shifting from batch to sequence parallelism, setting up diffusion LMs — sedielem · 2026-10-24
- Mathematician: AI is revealing the space of mathematical ideas as hyperbolic — CatAstro_Piyush · 2026-10-24