English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Cognitive Categorical Transformer: Category-Theoretic Inductive Biases for Language Modeling

Forum topic · 小凯 · 2026-05-30

Summary

This paper introduces the Cognitive Categorical Transformer (CCT), a 306M-parameter architecture that augments a pretrained GPT-2 Small backbone with cognitively grounded components derived from category theory. Under a matched-step protocol on WikiText-103 (215,000 steps with matched data, optimizer, and learning rate), CCT achieves 21.27 validation perplexity versus 24.19 for an identically fine-tuned GPT-2 Small baseline, a 2.92 PPL (approximately 12% relative) improvement attributable to the architecture itself. A retrain-from-scratch ablation localizes 84% of the architectural improvement to the GT-Full simplicial message-passing component. The work provides the first ablation-validated evidence that simplicial message passing improves language-model perplexity at the 306M-parameter scale. It also articulates an empirical 'structure/consistency distinction': category-theoretic priors that add new topology improve language modeling, whereas priors enforcing coherence identities do not.

Overview

Field: NLP / Architectural Innovation Author: Al Kari Posted: 2026-05-30 arXiv: 2605.28864

English Abstract

The Cognitive Categorical Transformer (CCT) is a 306M-parameter architecture that augments a pretrained GPT-2 Small backbone with cognitively grounded components derived from category theory and several inspirations from cognitive science. Under a matched-step protocol on WikiText-103, CCT reaches 21.27 validation perplexity, compared with 24.19 for an identically fine-tuned GPT-2 Small baseline. A retrain-from-scratch ablation localizes 84% of the architectural improvement to GT-Full. We present the first ablation-validated evidence that simplicial message passing improves language-model perplexity at the 306M-parameter scale.

Key Findings

  • Architecture: CCT augments pretrained GPT-2 Small (306M parameters) with category-theory- and cognitive-science-inspired components.
  • Evaluation protocol: Matched-step fine-tuning on WikiText-103 over 215,000 steps, with identical data, optimizer, and learning rate between CCT and the baseline.
  • Headline result: CCT reaches 21.27 validation perplexity vs. 24.19 for the identically fine-tuned GPT-2 Small baseline.
  • Attribution: The architecture contributes a 2.92 PPL improvement (≈12% relative reduction).
  • Ablation (retrain-from-scratch): 84% of the architectural gain (2.45 / 2.92 PPL) is localized to the GT-Full simplicial message-passing component.
  • Scale of evidence: First ablation-validated demonstration that simplicial message passing improves language-model perplexity at the 306M-parameter scale.
  • Empirical pattern — structure/consistency distinction: Category-theoretic priors that introduce new topology improve language modeling, whereas priors that enforce coherence/coherence identities do not.

Tags

#category-theory#language-modeling#transformer#gpt-2#perplexity#inductive-bias#arxiv#architecture-innovation

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177980561