MMODELYST
Papers/The Scaling Properties of Implicit Deductive Reasoning in Transformers
PAP

The Scaling Properties of Implicit Deductive Reasoning in Transformers

May 5, 2026

arXiv
Abstract

We investigate the scaling properties of implicit deductive reasoning over Horn clauses in depth-bounded Transformers. By systematically decorrelating provability from spurious features and enforcing algorithmic alignment, we find that in sufficiently deep models with a bidirectional prefix mask, implicit reasoning approaches explicit CoT performance across graph topologies and problem widths, though CoT remains necessary for depth extrapolation.

Select text to highlight · click a highlight to remove · saved in this browser only
Authors
Enrico Vompa, Tanel Tammet
Your notes (browser-local)
saved
arXiv:2605.04330