No Free Lunch GPT/LLM Limitations

Serge Berger

Keywords: first-order logic, 0-1 law, Borel probability space, GPT/LLM, CoT, hierarchical and recursive models

TL;DR: We present a first-order model that captures the reasoning limitations of modern deep learning architectures' inference (v.2)

Abstract: Recently, AI models (LLM/GPT) have demonstrated their reasoning capabilities in rigorous settings and their utility as research assistants. However, in this paper, we reveal a limitation of this approach in generating rigorous proofs (including with the help of RLHF). This limitation applies to (autoregressive) GPT/LLM-agentic and hierarchical models, as well as to a subset of EBM models.

Submission Number: 104