Understanding and Mitigating Dataset Corruption in LLM Steering

Research reveals that contrastive steering, a method for adjusting LLM behavior during inference, is moderately robust to data corruption but vulnerable to malicious attacks when significant portions of training data are compromised. The study identifies geometric patterns in corruption types and proposes using robust mean estimators as a safeguard against unwanted effects.

• →Contrastive steering shows resilience to moderate dataset corruption but fails when non-trivial fractions are maliciously altered.
• →Unwanted side effects can be clearly manifested through targeted corruption of training data used for steering directions.
• →The vulnerability stems from high-dimensional mean computation in the steering direction learning process.
• →Robust mean estimators can effectively mitigate most unwanted effects from malicious data corruption.
• →The research provides important insights for AI safety applications using contrastive steering methods.