Probabilistic coherence, logical consistency, and Bayesian learning: Neural language models as epistemic agents.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Computational Linguistics

Background:

• Neural language models (NLMs) are increasingly sophisticated but their capacity for rational thought is debated.
• Existing models may lack probabilistic coherence and logical consistency in their 'beliefs'.

Purpose of the Study:

• To investigate if suitably trained NLMs exhibit epistemic agency, characterized by rational belief revision.
• To introduce methods for measuring NLM belief coherence and rationality.

Main Methods:

• Computational experiments using T5 models (rankers) pretrained on synthetic corpora.
• Development of a procedure to elicit model beliefs and numerical metrics for rationality constraints (probabilistic, logical, Bayesian).
• Analysis of model performance before and after self-training on auto-generated texts.

Main Results:

• Pretrained rankers initially show global inconsistency.
• Self-training on synthetic data leads to probabilistically coherent and logically consistent belief systems.
• Self-training enables rational evidential learning by allowing belief propagation and adjustment.

Conclusions:

• The study confirms the Rationality Hypothesis: trained NLMs can exhibit advanced rational skills.
• Findings suggest NLMs possess capabilities extending beyond linguistic tasks, with broader implications.