Matching Human Consensus Feedback on Human Consensus Feedback
13.8PrecisionGOODPOINT-SFT
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| GOODPOINT-SFTBootstrap samples (B)=1000, Feedback per paper=52026.04 | 13.8 | 11.2 | 10.8 | |
| GPT-5.2Bootstrap samples (B)=1000, Feedback per paper=52026.04 | 13 | 16.5 | 13 | |
| Gemini-3-flashBootstrap samples (B)=1000, Feedback per paper=52026.04 | 12.8 | 16.9 | 13.1 | |
| GOODPOINT-DPOBootstrap samples (B)=1000, Feedback per paper=52026.04 | 9.3 | 10.7 | 8.7 | |
| Qwen3-8b (Base)Bootstrap samples (B)=1000, Feedback per paper=52026.04 | 6.9 | 8.4 | 6.8 | |
| Llama3.1-8b-InstructBootstrap samples (B)=1000, Feedback per paper=52026.04 | 4.7 | 5.3 | 4.4 |