Pairwise preference ranking on Held-out
1,187ELO ScoreHuman Abstracts
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Human Abstracts2025.10 | 1,187 | 17 | 1 | 94.4 | |
| GPT-OSS-120B + GPT-OSS-120BConfiguration=self-play2025.10 | 1,119 | 14 | 5 | 63.6 | |
| Mistral-24B + GPT-OSS-120BProposer=Mistral-24B, Verifier=GPT-OSS-120B2025.10 | 939 | 5 | 8 | 25 | |
| Mistral-24B + Mistral-24BConfiguration=self-play2025.10 | 927 | 5 | 12 | 25 | |
| GPT-OSS-120B + Mistral-24BProposer=GPT-OSS-120B, Verifier=Mistral-24B2025.10 | 828 | 1 | 16 | 5 |