Refusal-Gated Decoding: Preserving Refusal Behavior Under High-Temperature Sampling
The research addresses how increasing a Large Language Model's sampling temperature weakens its ability to refuse harmful prompts. The proposed solution is refusal-gated decoding, an efficient method that uses a quick initial check to see if the model will refuse before proceeding with slower high-temperature generation. This technique successfully maintains 91-99% of the original refusal behavior while keeping latency low.