Prefilling Attack

Improving LLM First-Token Predictions in\\Multiple-Choice Question Answering via Output Prefilling