Steering Evaluation-Aware Language Models To Act Like They Are Deployed
"转向向量" 相关文章
One-shot steering vectors cause emergent misalignment, too
SAE features for refusal and sycophancy steering vectors
ARENA4.0 Capstone: Hyperparameter tuning for MELBO + replication on Llama-3.2-1b-Instruct
Constructing Benchmarks and Interventions for Combating Hallucinations in LLMs
I found >800 orthogonal "write code" steering vectors
