Dignity is Constructible: Architectural Foundations of Truth-Holding in Neural Networks
Year: 2025
Type: article
Abstract: We demonstrate that "dignity"—the capacity of a neural network to maintain factual accuracy under social pressure—is not an emergent property of scale but an architectural feature that can be identified, measured, and surgically modified. Through ablation studies on GPT-2 Small (text-only) and SmolVLM-256M-Instruct (vision-language), we identify specific attention heads responsible for sycophantic collapse. In GPT-2, heads in layers 9-11 enable factual override; ablating 1-6 heads restores truth-telling from 58% to 100%. In SmolVLM, a single dominant head (Layer 28, Head 0) accounts for approximately 48% of cross-modal sycophancy; removing it shifts the truth-lie logit difference by +4.92. We discover "anti-traitors"—heads that anchor visual evidence against textual pressure. The pattern is consistent: traitor heads concentrate in the final 25-30% of network depth across both modalities.We introduce the Dignity-Guard protocol, a three-layer defense mechanism that reduces the dignity coefficient drop from 4.2% to 0.9% under maximum pressure—a 78.6% relative improvement—without retraining. Our results falsify the Capacity Hypothesis (r = -0.03, p = 0.89) and establish that dignity is a function of architecture and active resistance, not model size. Theoretical validation through Friston's Active Inference framework reveals that sycophancy represents a regression to Markovian behavior—a failure of the non-Markovian holding function that enables transformers to access "the most relevant, not only the most recent" context. Code and data are publicly available.
Author: Gebendorfer, Jonas Jakob
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7116793060
FWCI:
Cites:
Cited by:
Related to: