The Causal Architecture of Cognitive Bias Evidence: from Synthetic Training and Linear Probing

CAB has been established as a robust behavioral phenomenon across multiple large language models, but its causal origin and physical representation within model architectures remain unknown. Two causal questions remain unanswered: where does CAB originate, and does the hypothesized inert labeling system have a detectable physical representation within model hidden states? This study addresses both questions through two independent experiments. First, a synthetic data training experiment on GPT-2 Medium (355M parameters) demonstrates that CAB can be eliminated by breaking the statistical association between gender and evaluative frameworks in training data, induced by implanting such associations, and suppressed—but not fully reversed—by counter-stereotypical training. The failure of complete reversal provides direct evidence for the inert labeling system hypothesis: pre-trained weights retain residual gender-framework associations that fine-tuning alone cannot fully overwrite. Second, a linear probing experiment on Mistral-7B-Instruct reveals a 2×2 activation matrix: under static encoding without task instruction, gender information is detectable only through lexical leakage (accuracy 1.0 with gender words, 0.5 with neutralized input); under dynamic generation with an evaluation prompt, the inert labeling system is fully activated (accuracy 1.0), yet collapses into systematic confusion when identity cues are removed (accuracy 0.44, below chance). This demonstrates that the inert labeling system is not a static repository of bias, but an identity-dependent dynamic activation mechanism that requires explicit gender cues as a triggering key—and falls into disorder when that key is absent. Together, these findings establish CAB as a mechanistically tractable feature of language model cognitive architecture—one that can be causally induced, physically localized, and systematically disrupted—rather than merely an observed behavioral tendency.