OOD Generalization Tracking on Truthfulness datasets
0.994R^2 (MCS)Mahalanobis Cosine Similarity
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Mahalanobis Cosine SimilarityModel=Qwen2.5-7B2026.06 | 0.994 | — | |
| Mahalanobis Cosine SimilarityModel=Llama-70B, Layer=202026.06 | 0.99 | — | |
| Mahalanobis Cosine SimilarityModel=Llama-70B, Layer=332026.06 | 0.98 | — | |
| Mahalanobis Cosine SimilarityModel=Llama-70B, Layer=652026.06 | 0.976 | — | |
| Mahalanobis Cosine SimilarityModel=Llama-70B, Layer=502026.06 | 0.963 | — | |
| Mahalanobis Cosine SimilarityModel=Llama-3.1-8B2026.06 | 0.935 | — | |
| Euclidean Cosine SimilarityModel=Llama-70B, Layer=332026.06 | — | 0.441 | |
| Euclidean Cosine SimilarityModel=Llama-70B, Layer=202026.06 | — | 0.628 | |
| Euclidean Cosine SimilarityModel=Llama-70B, Layer=502026.06 | — | 0.422 | |
| Euclidean Cosine SimilarityModel=Llama-70B, Layer=652026.06 | — | 0.446 | |
| Euclidean Cosine SimilarityModel=Llama-3.1-8B2026.06 | — | 0.53 | |
| Euclidean Cosine SimilarityModel=Qwen2.5-7B2026.06 | — | 0.656 |