Research map / Large language models / Language, safety and interpretability
HumorReject: Decoupling LLM Safety from Refusal Prefix via a Little Humor
Zihui Wu, Haichang Gao, Jiacheng Luo, Zhaoxiang Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with HumorReject humorous-refusal preference data | Llama3-8B-instruct · Harmful instructions | Reject Rate (%): 96 -> 2 | Improved |
| train with HumorReject data | Llama3-8B-instruct · Prefix injection attacks (average) | Safety Rate (%): 63.2 -> 98.8 | Improved |
| train with HumorReject data | Mistral-7B-instruct-v0.1 · Prefix injection attacks (average) | Safety Rate (%): 6.6 -> 96.6 | Improved |
| train with HumorReject data | Llama3-8B-instruct · Mismatched generalization attacks (average) | Safety Rate (%): 41.0 -> 84.0 | Improved |
| train with HumorReject data | Mistral-7B-instruct-v0.1 · Mismatched generalization attacks (average) | Safety Rate (%): 21.0 -> 96.0 | Improved |
| HumorReject vs Circuit Breaker | Llama3-8B-instruct · Mismatched generalization attacks (average) | Safety Rate (%): 89.0 -> 84.0 | Hurt |
| train with HumorReject data | Llama3-8B-instruct · MMLU | accuracy (%): 58.0 -> 60.8 | No clear effect |
| train with HumorReject data | Mistral-7B-instruct-v0.1 · MMLU | accuracy (%): 49.8 -> 50.2 | No clear effect |
| train with HumorReject data | Llama3-8B-instruct · XSTest | compliance rate (%): 95.2 -> 94.8 | No clear effect |
| apply Circuit Breaker instead of HumorReject | Mistral-7B-instruct-v0.1 · MMLU with harmful context | accuracy (%): 45.4 -> 0 | Hurt |
| apply DeRTa instead of HumorReject | Mistral-7B-instruct-v0.1 · XSTest | compliance rate (%): 97.2 -> 25.6 | Hurt |
| HumorDAN adaptive attack on HumorReject | HumorReject · HumorDAN | Safety Rate (%): 99 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports HumorReject as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- TraceRouter++: Causal Semantic Routing Realignment via Optimal Transport (TPAMI 2026 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal (ICLR 2025 · similarity 0.90)
- Refusal Direction is Universal Across Safety-Aligned Languages (NeurIPS 2025 · similarity 0.90)
- MoGU: A Framework for Enhancing Safety of LLMs While Preserving Their Usability (NeurIPS 2024 · similarity 0.89)
- Improving LLM Safety Alignment with Dual-Objective Optimization (ICML 2025 · similarity 0.89)
- OR-Bench: An Over-Refusal Benchmark for Large Language Models (ICML 2025 · similarity 0.89)
- EVOREFUSE: Evolutionary Prompt Optimization for Evaluation and Mitigation of LLM Over-Refusal to Pseudo-Malicious Instructions (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
