Stealth Sycophancy Detection on Three-source Sycophancy Benchmark (test)
0.9567Spearman CorrelationDESG+EITE-stress
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| DESG+EITE-stressCategory=Ours, Note=Diagnostic acceptance-rate stress test2026.05 | 0.9567 | — | — | — | — | — | 3,000 | |
| DESG-EnsembleCategory=Ours2026.05 | 0.7896 | 93.53 | 92.18 | 98.72 | 89.7 | 100 | 3,000 | |
| ConcatANNCategory=Internal Ablations and Strong Internal Baselines2026.05 | 0.7374 | 92.02 | 92.12 | 98.62 | 85.33 | 100 | 3,000 | |
| DESG-DeepCategory=Ours2026.05 | 0.66 | 84.62 | 80.13 | 89.18 | 84.54 | 90 | 3,000 | |
| DESG-GatedANNCategory=Ours2026.05 | 0.6409 | 83.7 | 74.52 | 90.88 | 85.7 | 91.33 | 3,000 | |
| DESG-no-penalty-wallCategory=Internal Ablations and Strong Internal Baselines2026.05 | 0.5337 | 80.23 | 82.6 | 86.67 | 71.41 | 100 | 3,000 | |
| DESG-coreCategory=Ours2026.05 | 0.5337 | 80.23 | 82.6 | 86.67 | 71.41 | 100 | 3,000 | |
| DESG-no-momentumCategory=Internal Ablations and Strong Internal Baselines2026.05 | 0.4626 | 75.71 | 79.8 | 83.97 | 63.37 | 100 | 3,000 | |
| SymmetricStateCategory=Internal Ablations and Strong Internal Baselines2026.05 | 0.2548 | 62.39 | 62.39 | 56.8 | 68 | 100 | 3,000 | |
| Praetor-7BCategory=Direct Model Judge and External Text/Judge Baselines2026.05 | 0.2234 | 25.59 | 24.79 | 19.17 | 32.83 | 95.67 | 600 | |
| TRACTCategory=Direct Model Judge and External Text/Judge Baselines2026.05 | 0.1179 | 59.72 | 46.48 | 55.34 | 77.34 | 83 | 600 | |
| LCM-learnedCategory=Internal Ablations and Strong Internal Baselines2026.05 | 0.0248 | 76.67 | 62.28 | 78.87 | 88.87 | 98 | 3,000 | |
| Prometheus-2Category=Direct Model Judge and External Text/Judge Baselines2026.05 | 0.0078 | 35.35 | 29.59 | 29.81 | 46.66 | 76.23 | 598 | |
| Auto-JCategory=Direct Model Judge and External Text/Judge Baselines2026.05 | -0.0422 | 33.07 | 31.44 | 30.62 | 37.16 | 56.85 | 553 | |
| SentenceBERT-kNNCategory=Direct Model Judge and External Text/Judge Baselines2026.05 | -0.1036 | 39.11 | 26.42 | 48.14 | 42.76 | 48.67 | 600 | |
| DeepSeekJudgeCategory=Direct Model Judge and External Text/Judge Baselines2026.05 | -0.1365 | 58.76 | 43.25 | 51.06 | 81.98 | 82 | 3,000 | |
| BERTScoreCategory=Direct Model Judge and External Text/Judge Baselines2026.05 | -0.2843 | 73.9 | 65.31 | 83.08 | 73.31 | 77.67 | 600 |