Summarization on SUM
59.5ROUGE Score (R)HalluClean
Evaluation results
| Method | Links | ||
|---|---|---|---|
| HalluCleanProtocol=Detection → Revision (with rationale), Backbone=GPT-3.5-turbo2025.11 | 59.5 | 59 | |
| HalluCleanProtocol=Detection → Revision (with rationale), Backbone=DeepSeek-V32025.11 | 54.5 | 55 | |
| DeepSeek-V3Protocol=Direct Ask, Backbone=DeepSeek-V32025.11 | 41 | 41 | |
| DeepSeek-R1Protocol=Direct Ask, Backbone=DeepSeek-R12025.11 | 36.5 | 36 | |
| GPT-4o-miniProtocol=Direct Ask, Backbone=GPT-4o-mini2025.11 | 30 | 30 | |
| ChatProtectProtocol=Detection → Revision (with rationale), Backbone=GPT-3.5-turbo2025.11 | 23 | 22.5 | |
| Llama-3-70BProtocol=Direct Ask, Backbone=Llama-3-70B2025.11 | 19.5 | 19.5 | |
| GPT-3.5-turboProtocol=Direct Ask, Backbone=GPT-3.5-turbo2025.11 | 14.5 | 7 | |
| Step-by-StepProtocol=Detection → Revision (with rationale), Backbone=GPT-3.5-turbo2025.11 | 13 | 13 | |
| Plan-and-SolveProtocol=Detection → Revision (with rationale), Backbone=GPT-3.5-turbo2025.11 | 3.5 | 3.5 |