Over-Refusal
Refers to the behavior of large language models refusing harmless requests that should normally be answered. Together with safety rate, it forms a tension metric that jointly characterizes the safety-usefulness frontier.
Timeline (last 90 days)
When fine-tuned on samples of safe refusals, the model learns refusal as a tendency, which may lead it to over-refuse legitimate questions as well.
In the controlled sweep of the general-data proportion in safety SFT, the over-refusal curve is non-monotonic, with the inflection point falling in the 40%–50% general-data proportion range.
All Facts (2)
When fine-tuned on samples of safe refusals, the model learns refusal as a tendency, which may lead it to over-refuse legitimate questions as well.
50%UnverifiedIn the controlled sweep of the general-data proportion in safety SFT, the over-refusal curve is non-monotonic, with the inflection point falling in the 40%–50% general-data proportion range.
50%