Tag
This study investigates off-target effects of response-style alignment in a Korean 27B language model, finding that post-training for style significantly impacts answer propensity and disclosure rates without targeting safety or capability.