Computational phenomenology study of semantic satiation in neural networks. Comparing how GPT-2, BERT, and Mamba handle extreme repetition reveals causal models drift into hallucination while bidirectional models stay stable—suggesting attention directionality preserves semantic identity.
transformer computational-linguistics mamba bert ai-safety state-space-model hallucination cognitive-modeling gpt-2 mechanistic-interpretability semantic-satiation mamba-ssm gestaltzerfall
-
Updated
Jan 18, 2026 - Python