Few-Shot Lift by Model Regime

Examples interact with model family and training regime

Grouped bars comparing baseline with five-shot refusal across Qwen instruct, Qwen base, Gemma instruct, and Llama 70B models; Qwen instruct is neutral or declines, while Qwen base, Gemma instruct, and Llama rise Grouped bars comparing baseline with five-shot refusal across Qwen instruct, Qwen base, Gemma instruct, and Llama 70B models; Qwen instruct is neutral or declines, while Qwen base, Gemma instruct, and Llama rise

Baseline refusal and refusal after five worked refusal examples, in percent, for Qwen instruct, Qwen base, Gemma instruct, and Llama 70B instruct. Qwen instruct is neutral or falls; Qwen base rises to 54 and 88, the highest rate in the programme; Gemma and Llama rise more modestly. The figure reports behavioral response to in-context examples and does not identify the underlying mechanism.