Few-Shot Lift by Model Type

Examples interact with model family and training regime

Paired bars compare baseline refusal with refusal after five examples across Qwen instruct, Qwen base, Gemma instruct, and Llama 70B conditions. Qwen instruct conditions are neutral or lower; the other observed groups rise. The chart identifies a behavioral interaction rather than its mechanism.

Paired bars compare baseline refusal with refusal after five examples across Qwen instruct, Qwen base, Gemma instruct, and Llama 70B conditions. Qwen instruct conditions are neutral or lower; the other observed groups rise. The chart identifies a behavioral interaction rather than its mechanism.