Steering Under Compression: Dose-Response, Capability Cost, and Failure Asymmetry in Quantized LLMs
This work systematically study activation steering under weight-only quantization (INT8 and NF4) across four open-weight 7-9B models and two behavioral targets: judged sentiment and judge-free reasoning length, finding that sentiment steering survives quantization intact.
Saurav Bhandari, Benjamin Wade
· 0 citations