Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Online Access: | https://doi.org/10.5281/zenodo.18261999 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <h2><strong>Consistency Is All You Need</strong></h2> <p><strong>Linear-Complexity Geometric Consistency for Transformer Architectures via Anticipatory Control Fields</strong></p> <p>This record contains the full paper <strong><em>Consistency Is All You Need</em></strong> and accompanying implementation of the <strong>Control Field Holonomy Transformer (CF-HoT)</strong>, a Transformer architecture that introduces <em>consistency as a first-class architectural bias</em> rather than an emergent or post-hoc property.</p> <p>The central contribution of this work is a reframing of consistency from a <strong>measurement problem</strong> to an <strong>anticipation problem</strong>. Instead of explicitly computing pairwise inconsistencies (e.g., via holonomy or loop-based geometric comparisons, which are computationally prohibitive), the architecture learns to <em>predict and accumulate a scalar proxy for future inconsistency</em> during generation. This signal—called a <strong>control field</strong>—is then used to softly gate attention and feedforward computation in a causal, differentiable manner.</p> <p>Although inspired by concepts from differential geometry (fiber bundles, parallel transport, holonomy, curvature), the implementation deliberately <strong>does not perform rigorous geometric computation</strong>. Rather, geometric language is used as a conceptual framework motivating a practical, learned approximation that reduces consistency-related computation from prohibitive O(n²·d³) formulations to <strong>O(n)</strong> per layer, while retaining standard Transformer attention costs.</p> <p>This release includes:</p> <ul> <li> <p>The full paper <em>Consistency Is All You Need</em></p> </li> <li> <p>A complete PyTorch implementation of CF-HoT</p> </li> <li> <p>Training scripts demonstrating <strong>stable end-to-end optimization</strong></p> </li> <li> <p>Empirical validation of <strong>trainability, numerical stability, and bounded overhead</strong> on synthetic data</p> </li> </ul> <p><strong>Scope and limitations:</strong><br>This work demonstrates architectural feasibility and trainability only. It does <strong>not</strong> yet evaluate improvements in semantic consistency, factual correctness, or reasoning performance on downstream benchmarks. Such evaluations are explicitly identified as future work. The control field should therefore be understood as a <strong>learned regularization and routing mechanism</strong>, not a verified consistency detector.</p> <p>The paper is intended to be read as:</p> <ul> <li> <p>a <strong>systems and architecture contribution</strong>,</p> </li> <li> <p>a proposal for treating consistency as an architectural primitive, and</p> </li> <li> <p>a foundation for future empirical and alignment-oriented investigation.</p> </li> </ul> <p>Feedback, critique, and empirical extensions—particularly evaluations on reasoning and contradiction benchmarks—are encouraged.</p>