Logical Reasoning on ProofW (Accuracy, Token Cost)
77.5AccuracyDenser
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DenserBackbone=Qwen3-14B2025.12 | 77.5 | -58.7 | |
| Process SupervisionBackbone=Qwen3-14B2025.12 | 76.8 | 156.3 | |
| Self-VerificationBackbone=Qwen3-14B2025.12 | 76.2 | 142.8 | |
| Reflection-CoTBackbone=Qwen3-14B2025.12 | 75.3 | 127.6 | |
| TokenSkip (0.9 ratio)Backbone=Qwen3-14B2025.12 | 75.1 | -53.5 | |
| Tree-of-ThoughtBackbone=Qwen3-14B2025.12 | 74.9 | 345.7 | |
| Think-to-ThinkBackbone=Qwen3-14B2025.12 | 74.2 | 187.2 | |
| Self-ConsistencyBackbone=Qwen3-14B2025.12 | 73.6 | 289.4 | |
| Chain-of-Thought (CoT)Backbone=Qwen3-14B2025.12 | 72.1 | 0 | |
| BeConciseBackbone=Qwen3-14B2025.12 | 70.4 | -54.2 | |
| AbbreWordsBackbone=Qwen3-14B2025.12 | 69.8 | -50.8 |