A Logic for Expressing Log-Precision Transformers

Authors: William Merrill, Ashish Sabharwal

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical We prove any log-precision transformer classifier can be equivalently expressed as a first-order logic sentence that, in addition to standard universal and existential quantifiers, may also contain majority-vote quantifiers. This is the tightest known upper bound and first logical characterization of log-precision transformers.
Researcher Affiliation Collaboration William Merrill New York University willm@nyu.edu Ashish Sabharwal Allen Institute for AI ashishs@allenai.org
Pseudocode Yes Algorithm 1 node C(n, i) Return the type of gate i in circuit Cn. Algorithm 2 edge C(n, i, j) If Cn contains an edge i ! j, return the argument number of that edge. Otherwise, return 1.
Open Source Code No The paper does not mention any open-source code for the methodology described.
Open Datasets No The paper does not discuss or provide access information for any dataset used for training, as it is a theoretical paper.
Dataset Splits No The paper does not provide information about training/validation/test dataset splits, as it is a theoretical work without empirical evaluations.
Hardware Specification No The paper does not provide specific hardware details, as it is a theoretical work without experimental setup.
Software Dependencies No The paper does not provide specific software dependency details with version numbers, as it is a theoretical work without experimental setup.
Experiment Setup No The paper does not contain specific experimental setup details (e.g., hyperparameter values), as it is a theoretical work without an empirical setup.