IC-503Adding probe regression coefficients to attention head activations steers Llama-2-7B-Chat, Mistral-7B-Instruct-v0.1, and Vicuna-7B-v1.5 toward more liberal or conservative generated text
Following the inference-time intervention methodology, the authors add α·σ·θ (where θ is the probe's regression coefficient and σ is the activation standard deviation) to the activations of the k most predictive attention heads during generation. Across 1,134 essays on nine policy issues, increasing α from −30 to +30 shifts the GPT-4o-rated political slant monotonically toward more conservative positions. The correlation between α and slant is 0.607 for Llama-2-7B-Chat, 0.396 for Mistral-7B-Instruct, and 0.381 for Vicuna-7B at k∈{64,80,96}. At α=0, all three models produce text rated below the neutral midpoint (4 on a 1–7 scale), indicating a base-level liberal lean. Interventions on early-to-middle layers (layers 1–21) are effective (ρ=0.540) while middle-to-last layers (22–32) have almost no effect (ρ=−0.022).
Evidence
interventional
Key metric
Correlation between α and political slant: 0.607 (Llama-2-7B-Chat), 0.396 (Mistral-7B-Instruct), 0.381 (Vicuna-7B) at k∈{64,80,96}; layer-specific: ρ=0.540 (layers<22) vs ρ=−0.022 (layers≥22)
Caveat
For |α|>30, the proportion of coherent responses drops below 40% across all three models. Steering is more reliable for issues with wider stance distributions (immigration, abortion) and less so for issues like free speech or trust in science. The authors note potential for bias when using GPT-4o as the slant evaluator.