Mathematical Reasoning on GSM8K (n=300)
68.8Accuracy @73.7%Max positive ΔH
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Max positive ΔHCoverage=73.7%2026.03 | 68.8 | 5.8 | 0.135 | 0.34 | |
| Violation count (−vc)Coverage=73.7%2026.03 | 68.8 | 5.8 | 0.209 | 0.311 | |
| Entropy monotonicity (ours)Coverage=73.7%2026.03 | 68.8 | 5.8 | 0.2 | 0.311 | |
| Chain length (shorter first)Coverage=73.7%2026.03 | 65.6 | 2.6 | 0.101 | 0.31 | |
| Final-step entropy HNCoverage=73.7%2026.03 | 65.2 | 2.2 | 0.093 | 0.313 | |
| Yes/No self-eval (P(Yes))Coverage=73.7%2026.03 | 63.3 | 0.3 | 0.019 | 0.395 | |
| RandomCoverage=73.7%2026.03 | 62.9 | 0.1 | 0 | 0.398 | |
| Scalar coherence H0−HNCoverage=73.7%2026.03 | 62.4 | 0.6 | 0.059 | 0.408 | |
| Self-judgment (1 verifier call)Coverage=73.7%2026.03 | 62.4 | 0.6 | 0.019 | 0.368 |