Do LLMs Act Like Rational Agents? Measuring Belief Coherence in Probabilistic Decision Making

Khurram Yamin, Jingjing Tang, Santiago Cortes-Gomez, Amit Sharma, Eric Horvitz, Bryan Wilder · 2026

Read Do LLMs Act Like Rational Agents? Measuring Belief Coherence in Probabilistic Decision Making by Khurram Yamin, Jingjing Tang, Santiago Cortes-Gomez, Amit Sharma, Eric Horvitz, Bryan Wilder (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-06286.txt.

Description

When language models state a probability, do their choices match what they claim to believe? Khurram Yamin and 5 co-authors build a decision-theory test for that consistency. The framework asks for both a probability judgment and a decision, then checks whether a near-rational decision maker holding that belief could have made that choice.

Surprisingly, the test needs no assumption about what the agent values. Applied to stylized medical diagnosis tasks, the method found that reported beliefs are imperfect summaries of the information inside model decisions. The gaps were small for the strongest models, though. The work offers a way to validate machine beliefs before trusting them in high-stakes settings.

Catalog record

CREATORKhurram Yamin, Jingjing Tang, Santiago Cortes-Gomez, Amit Sharma, Eric Horvitz, Bryan Wilder
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-06286

Browse the full library · About the Tower of Babel