A single linear partisan axis at layer 18 of Llama 3.1 8B Instruct separates Republican from Democratic congressional text (AUC 0.945, Cohen's d 1.94), and ablating or amplifying the partisan component mid-generation systematically shifts output

- Tam trains per-layer logistic-regression probes on Llama 3.1 8B Instruct residual-stream activations using 190,491 tweets from members of the U.S. Congress; at layer 18 the probe's learned weight vector separates Republican from Democratic text with a projection AUC of 0.945 and Cohen's d of 1.94 (cross-validated probe AUC 0.935), and a TopK sparse autoencoder (k=64) is then used to decompose the axis into partly interpretable partisan features. A simple mean-difference direction at the same layer reaches an AUC of only 0.715, which the author reads as suggesting the partisan representation is multi-dimensional. [tam-2026-amplifying-mirror-partisan-direction] - Ablating or amplifying the probe-direction component at layer 18 (fractions 0.15-0.3 of the hidden-state norm, 150 prompts) produces systematic shifts in output (stance inversion, register shifting, named-entity fabrication), shown through qualitative examples; ablation is uneven and leaves a residual lean on some prompts. This is evidence that the axis is a causally effective direction, but it is measured on a single model and a single labeled corpus without a quantitative steering metric, so it should be read as directional rather than general. [tam-2026-amplifying-mirror-partisan-direction]