Step-wise Verification on ProcessBench GSM8K v1 (val)
68.6True Negative RateLlama-3.3-70B
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Llama-3.3-70BBackbone=Llama-3.3-70B, Decoding Strategy=Greedy2026.05 | 68.6 | 96.9 | 80.3 | |
| VerifySteer-BiBackbone=Qwen3-8B, Decoding Strategy=Greedy2026.05 | 64.3 | 93.8 | 76.3 | |
| Cons@4Backbone=Qwen3-8B, Decoding Strategy=Sampling (T=0.7, p=0.8)2026.05 | 63.3 | 95.6 | 76.2 | |
| VerifySteer-UniBackbone=Qwen3-8B, Decoding Strategy=Greedy2026.05 | 62.3 | 93.3 | 74.7 | |
| Qwen2.5-72BBackbone=Qwen2.5-72B, Decoding Strategy=Greedy2026.05 | 61.8 | 95.9 | 75.2 | |
| Qwen3-14BBackbone=Qwen3-14B, Decoding Strategy=Greedy2026.05 | 61.4 | 98.4 | 75.6 | |
| Prompt Opt.Backbone=Qwen3-8B, Decoding Strategy=Greedy2026.05 | 58.5 | 94.3 | 72.2 | |
| Basic PromptBackbone=Qwen3-8B, Decoding Strategy=Greedy2026.05 | 54.6 | 96.4 | 69.7 | |
| VerifySteer-BiBackbone=Qwen3-1.7B, Decoding Strategy=Greedy2026.05 | 27.1 | 83.4 | 40.9 | |
| VerifySteer-UniBackbone=Qwen3-1.7B, Decoding Strategy=Greedy2026.05 | 25.6 | 77.2 | 38.5 | |
| Prompt Opt.Backbone=Qwen3-1.7B, Decoding Strategy=Greedy2026.05 | 22.2 | 76.7 | 34.5 | |
| Cons@4Backbone=Qwen3-1.7B, Decoding Strategy=Sampling (T=0.7, p=0.8)2026.05 | 21.6 | 81.2 | 34.1 | |
| Basic PromptBackbone=Qwen3-1.7B, Decoding Strategy=Greedy2026.05 | 17.9 | 96.4 | 30.2 |