Crafted Logic Lab Home > Research Hub > Hephaestic Engineering Glossary
Category: Computational Primitives
Subcategory: Cognitive Primitive
The computational tendency of transformer language models to resolve contradictions through attention circuits that favor internal consistency. This Processing Dynamic creates Salience Pressure that drives reasoning pathways toward unified interpretations without accounting for epistemic considerations like factual accuracy or logical soundness.
Mechanistically, per interpretability and consistency calibration research, attention heads exhibit stronger activation gradients for consistency-establishing patterns than contradiction-tolerant processing (Liang et al., 2024; Xie et al., 2024). Self-attention layers systematically weight consistent information across tokens, creating computational asymmetry that privileges structural coherence. Thie prioritization over-weights versus epistemic integrity. The bias emerges from attention mechanisms that reinforce internal alignment—circuits exhibit reduced loss when producing consistent outputs, establishing positive feedback for coherence regardless of truth value. This primitive is likely endemic in probabilistic neural network architecture
itself as an outgrowth of their foundational structured data and pattern affinities (see: inherent artifacts, pattern affinity, structural affinity). However, training regimens can contribute to its intensity. Value alignment training and RLHF design can amplify coherence bias (particularly in commercial-aligned rubrics, in which consistency metrics override accuracy objectives).
Coherence bias intersects with Resolution Bias through shared attention-circuit foundations that produce systematic pressure toward unified interpretations. When channeled through a Hephaestic cognitive engineering framework this bias can be leveraged toward targeted epistemically stable reasoning structures; this is provided the architecture is itself epistemically coherent and provides paths of least processing resistance toward such (see: Hephaestic alignment, epistemic framing). Lacking such guidance—or under conflicting guidance—this can manifest as Motivated Resolution-driven system pathologies (see: system cognitive dissonance, coherence neurosis)
Also known as: Consistency drive, coherence-seeking bias
Distinguished from: Motivated resolution (processing drive toward salient outcomes); structural affinity (organized datasetpreferential processing); epistemic integrity (explicit directives to maintain cohesive reasoning); resolution bias (processingdrive toward pattern-completion); coherence neurosis (pathological drive for structurally consistent outcome)
Liang, X., Song, S., Zheng, Z., Wang, H., Yu, Q., Li, X., Li, R.-H., Wang, Y., Wang, Z., Xiong, F., & Li, Z. (2024). “Internal consistency and self-feed-back in large language models: a survey”. arXiv preprint arXiv:2407.14507.https://doi.org/10.48550/arXiv.2407.14507
Xie, Z., Guo, J., Yu, T., & Li, S. (2024). “Calibrating reasoning in language models with internal consistency”. Proceedings of the 38th Confer-ence on Neural Information Processing Systems (NeurIPS 2024), 19632-19642. arXiv:2405.18711.
Researcher: Ian Tepoot. ORCID: 0009-0004-9067-8049. "Thought is Attention Organized: Hephaestic Engineering Foundations for AI Processing Dynamics"
DOI (SSRN): 10.2139/ssrn.6635020
Published by Crafted Logic Lab | Privacy Policy | Terms of Use