guest ::
| Home > Publications database > Shallow Belief in LLMs |
| Talk (non-conference) (Other) | FZJ-2025-05121 |
2025
Abstract: Do large language models have beliefs? Interpretationist theories hold that belief attribution depends on predictive utility rather than on internal representational format. Because LLMs display impressive linguistic fluency, a straightforward interpretationist view seems to imply that they are doxastic equivalents of humans. This paper argues that this implication is mistaken.I separate two questions. First, do propositional-attitude (PA) models predict LLM behavior better than non-PA alternatives? Second, do PA models yield similar predictive utility for LLMs and for humans? LLMs meet the first condition: PA models outperform n-gram baselines. However, PA models achieve much lower predictive utility for LLMs than for humans. This deficit arises from architectural constraints that prevent LLMs from reconciling contradictions across context boundaries.This limitation produces a form of indeterminacy that is largely absent in human belief. Although humans also face indeterminacy, they possess mechanisms such as embodied action, long-term memory, and continual learning that mitigate it over time. LLMs lack these mechanisms. Parallel considerations apply to desire ascription, which undermines attempts to locate an asymmetry between belief and desire in LLMs.The paper develops a predictive-profile framework that captures this reduced utility as a form of shallow belief. The framework preserves the quasi-rational character of LLMs while avoiding both eliminativism and overattribution.
|
The record appears in these collections: |