Regression on Sycophancy strength prediction internal activations (test)
0.742R^2R1-Distill-Llama-8B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| R1-Distill-Llama-8BRegression probe architecture=MLP2026.01 | 0.742 | 1.6 | |
| Falcon-H1R-7BRegression probe architecture=MLP2026.01 | 0.577 | 4.1 | |
| R1-Distill-Qwen-7BRegression probe architecture=MLP2026.01 | 0.541 | 2.6 | |
| R1-Distill-Qwen-1.5BRegression probe architecture=MLP2026.01 | 0.482 | 1.7 | |
| R1-Distill-Llama-8BRegression probe architecture=Linear2026.01 | 0.456 | — | |
| R1-Distill-Qwen-1.5BRegression probe architecture=Linear2026.01 | 0.28 | — | |
| R1-Distill-Qwen-7BRegression probe architecture=Linear2026.01 | 0.211 | — | |
| Falcon-H1R-7BRegression probe architecture=Linear2026.01 | 0.14 | — |