Hallucination Detection on Math (test)
98.8AuROCmarginalized energy
Evaluation Results
| Method | Links | |
|---|---|---|
| marginalized energyLLM=Mistral, Aggregation Strategy=Min, Notation=Marginal E^m2026.02 | 98.8 | |
| spilled energyLLM=Mistral-Instruct, Aggregation Strategy=Min, Notation=Spilled ΔE2026.02 | 95.95 | |
| Spilled ΔEModel=Gemma-Instruct 4B, Pool=Min2026.02 | 82.29 | |
| Logit EℓModel=Gemma-Instruct 4B, Pool=Max2026.02 | 78.64 | |
| Marginal EmModel=Gemma-Instruct 4B, Pool=Min2026.02 | 76.03 | |
| Marginal EmModel=Gemma-Instruct 4B, Pool=Max2026.02 | 74.57 | |
| Logit EℓModel=Gemma-Instruct 1B, Pool=Max2026.02 | 65.12 | |
| Marginal EmModel=Gemma-Instruct 1B, Pool=Max2026.02 | 64.21 | |
| Marginal EmModel=Gemma-Instruct 1B, Pool=Min2026.02 | 63.2 | |
| Spilled ΔEModel=Gemma-Instruct 1B, Pool=Min2026.02 | 56.82 | |
| Spilled ΔEsModel=Gemma-Instruct 4B, Pool=Max2026.02 | 43.94 | |
| Spilled ΔEsModel=Gemma-Instruct 1B, Pool=Max2026.02 | 40.49 |