| title | The Scaling Properties of Implicit Deductive Reasoning in Transformers |
|---|---|
| authors | Enrico Vompa, Tanel Tammet |
| date | 2026-05-05 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2605.04330 |
| affiliation | Applied AI group of the Tallinn University of Technology |
| upvotes | 3 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf |
| 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. |
cite
@misc{vompa2026scaling,
title={The Scaling Properties of Implicit Deductive Reasoning in Transformers},
author={Enrico Vompa and Tanel Tammet},
year={2026},
eprint={2605.04330},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2605.04330}
}