General Chat on ArenaHard v2.0
52Win RateGOLF
Evaluation Results
| Method | Links | |
|---|---|---|
| GOLFModel Backbone=Qwen-3-8B2026.03 | 52 | |
| Pairwise-GRPOModel Backbone=Qwen-3-8B2026.03 | 50.1 | |
| Rubric-as-RewardModel Backbone=Qwen-3-8B2026.03 | 50.08 | |
| Critique-GRPOModel Backbone=Qwen-3-8B2026.03 | 49.63 | |
| Direct-LikertModel Backbone=Qwen-3-8B2026.03 | 41.7 | |
| Qwen-3-8BModel Backbone=Qwen-3-8B2026.03 | 33.9 | |
| GOLFModel Backbone=Llama-3.1-8B-Instruct2026.03 | 25.03 | |
| Rubric-as-RewardModel Backbone=Llama-3.1-8B-Instruct2026.03 | 15.57 | |
| Critique-GRPOModel Backbone=Llama-3.1-8B-Instruct2026.03 | 13.73 | |
| Pairwise-GRPOModel Backbone=Llama-3.1-8B-Instruct2026.03 | 13.3 | |
| Direct-LikertModel Backbone=Llama-3.1-8B-Instruct2026.03 | 11.73 | |
| Llama-3.1-8B-InstructModel Backbone=Llama-3.1-8B-Instruct2026.03 | 5.57 |