Model refuses without warning
The study reveals that model refusal is not token prediction but the execution of explicit instructions establishing boundaries.
- Field
- training language models
- What they did
- Researchers studied the phenomenon where a language model refuses to answer a question by not generating a predicted token, but simply remaining silent.
- Why it matters
- This is important because such refusals show boundaries where the model's statistical machinery encounters explicit instructions it cannot ignore, unlike normal text prediction.
#llm#refusal#alignment#safety



