Can Transformers Reason Logically? A Study in SAT Solving

We theoretically and empirically study the logical reasoning capabilities of LLMs in the context of the Boolean satisfiability (SAT) problem. First, we construct a decoder-only Transformer that can solve SAT using backtracking and deduction via Chain-of-Thought (CoT). We prove its correctness by showing trace equivalence to the well-known DPLL SAT-solving algorithm. Second, to support the implementation of this abstract construction, we design a compiler $\texttt{PARAT}$ that takes as input a procedural specification and outputs a transformer model implementing this specification. Third, rather than $\textit{programming}$ a transformer to reason, we evaluate empirically whether it can be $\textit{trained}$ to do so by learning directly from algorithmic traces ("reasoning paths") of the DPLL algorithm.

Towards Understanding Neural Collapse: The Effects of Batch Normalization and Weight Decay

Neural Collapse (NC) is a geometric structure recently observed in the final layer of neural network classifiers. In this paper, we investigate the interrelationships between batch normalization (BN), weight decay, and proximity to the NC structure. Our work introduces the geometrically intuitive intra-class and inter-class cosine similarity measure, which encapsulates multiple core aspects of NC. Leveraging this measure, we establish theoretical guarantees for the emergence of NC under the influence of last-layer BN and weight decay, specifically in scenarios where the regularized cross-entropy loss is near-optimal. Experimental evidence substantiates our theoretical findings, revealing a pronounced occurrence of NC in models incorporating BN and appropriate weight-decay values. This combination of theoretical and empirical insights suggests a greatly influential role of BN and weight decay in the emergence of NC.