Recognition–action coupling in Qwen 2.5 7B
Saved-score contrast, pending fold-local preprocessing validation
Saved probe-score correlations (Spearman ρ, within 182 adversarial prompts, Qwen 2.5 7B layer 27): base ρ = −0.270, instruct +0.036, bilateral +0.458. The bilateral-minus-instruct contrast is +0.421, with a saved-score paired-bootstrap interval of [+0.281, +0.554]. The scaler and PCA were fitted before the folds, allowing held-out examples to influence preprocessing. These historical correlations remain descriptive; inductive validation requires a fold-local refit. The shaded band and intervals summarize the saved scores.