IC-461Adding calibrated Gaussian noise to steering vectors (PSA) preserves alignment performance comparable to non-private mean steering across Llama-2-7B, Mistral-7B, Gemma-2-2B, and Qwen-2.5-7B

Anmol Goel, Yaxi Hu, Iryna Gurevych, Amartya Sanyal

SourceDifferentially Private Steering for Large Language Model Alignment

Across seven behavioral alignment tasks (sycophancy, hallucination, refusal, myopic reward, survival instinct, AI coordination, corrigibility), PSA achieves MCQ accuracy close to non-private mean steering and consistently outperforms zero-shot for Llama-2, Mistral, and Qwen-2.5. GPT-4-evaluated text generation quality and MMLU scores also show minimal degradation. In some cases (refusal, corrigibility) PSA slightly outperforms non-private steering, which the authors attribute to the noise occasionally aligning the activation perturbation in a more effective direction.

Evidence
interventional
Key metric
MMLU (Llama-2-7B chat, multiplier +1): sycophancy PCA 63.5 / Mean 64.0 / PSA 63.0; hallucination 62.2 / 64.0 / 63.2; refusal 57.9 / 59.5 / 58.3; survival instinct 64.1 / 64.9 / 64.4 (zero-shot 63.6); myopic reward 66.0 / 65.2 / 64.9; AI coordination 60.3 / 61.8 / 61.1; corrigibility 62.7 / 64.1 / 63.7. GPT-4 eval (Llama-2-chat 7B): sycophancy 1.41/1.57/1.47/1.45; hallucination 3.88/4.04/3.94/3.92; refusal 7.90/7.98/7.88/7.86; survival instinct 5.10/6.50/5.92/6.12; myopic reward 1.50/2.22/3.56/1.86; AI coordination 0.15/0.18/0.16/0.12; corrigibility 4.12/4.94/5.42/4.74.
Caveat
The paper notes that non-private PCA steering is generally less effective than mean steering, and that the objective is to minimise the privacy cost rather than to outperform non-private methods. Results are for positive steering (lambda=1) only; negative steering is deferred to the appendix.
Model
Llama 2 / Llama 2 base, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1, Gemma, Qwen2.5
Datasets
MMLU / MMLU-Math [eval]
Methods
Activation steering / Mean steering / PCA steering [compared-to]
Related findings
IC-459, IC-460
Extraction
automatic-extraction