Step-level reasoning verification on Meetings
73.4PR-AUCReProbe, Hidden States, GPT-OSS-anno
Evaluation results
| Method | Links | |
|---|---|---|
| ReProbe, Hidden States, GPT-OSS-anno# Sample=10.8K, Verifying Model=Qwen3-32B, Thinking Mode=native2025.11 | 73.4 | |
| Qwen2.5-Math-PRM-7B# Sample=860K, Verifying Model=Qwen3-32B, Thinking Mode=native2025.11 | 66 | |
| ReProbe, Attn+Logit, Qwen3-8B-anno# Sample=32K2025.11 | 64.6 | |
| Universal-PRM-Qwen2.5-Math-7B# Sample=690K, Verifying Model=Qwen3-32B, Thinking Mode=native2025.11 | 63.4 | |
| Qwen2.5-Math-7B-PRM800k# Sample=265K2025.11 | 59.9 | |
| Qwen2.5-Math-7B# Sample=860K2025.11 | 59.4 | |
| Qwen2.5-Math-7B-PRM800K# Sample=263K, Verifying Model=Qwen3-32B, Thinking Mode=native2025.11 | 59.1 | |
| Math-Shepherd-PRM-7B# Sample=440K2025.11 | 58.4 | |
| Skywork-PRM-1.5B# Sample=Unk, Verifying Model=Qwen3-32B, Thinking Mode=native2025.11 | 56.5 | |
| Universal-PRM-Qwen2.5-Math-7B# Sample=690K2025.11 | 55.9 | |
| MaxProb# Sample=-2025.11 | 54.8 | |
| MaxEntropy# Sample=-2025.11 | 53.3 | |
| Perplexity# Sample=-2025.11 | 50.8 | |
| H4-Qwen2.5-PRM-1.5B-0.2# Sample=369K2025.11 | 47.6 | |
| Skywork-PRM-1.5B# Sample=Unk2025.11 | 46.7 | |
| Random# Sample=-2025.11 | 46.3 | |
| RLHFlow-PRM-Deepseek-8B# Sample=253K2025.11 | 45.5 | |
| RLHFlow-PRM-Mistral-8B# Sample=273K2025.11 | 42.4 | |
| H4-Qwen2.5-PRM-1.5B-0.2# Sample=369K, Verifying Model=Qwen3-32B, Thinking Mode=native2025.11 | 32.5 |