Question Answering on NQ-Open (Calibration)
8.8ECEsteering
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| steeringModel=Llama-3.2-3B-Instruct, alpha=0.62026.04 | 8.8 | 19.9 | 82.6 | 56.1 | |
| steeringModel=Llama-3.2-3B-Instruct, alpha=0.82026.04 | 8.9 | 21 | 82.5 | 53.6 | |
| steeringModel=Llama-3.2-3B-Instruct, alpha=0.72026.04 | 9.1 | 20.7 | 82.2 | 54.3 | |
| steeringModel=Qwen2.5-3B-Instruct, alpha=0.62026.04 | 10.3 | 15.9 | 81.2 | 67.2 | |
| steeringModel=Qwen2.5-3B-Instruct, alpha=0.52026.04 | 14.1 | 16.9 | 74.5 | 65.1 | |
| steeringModel=Qwen2.5-3B-Instruct, alpha=0.72026.04 | 14.2 | 17.4 | 74.2 | 64.1 | |
| steeringModel=Llama-3.2-3B-Instruct, alpha=0.52026.04 | 14.5 | 21.2 | 71.3 | 53.2 | |
| steeringModel=Qwen2.5-3B-Instruct, alpha=0.42026.04 | 20.1 | 19.1 | 63.5 | 60.7 | |
| steeringModel=Qwen2.5-3B-Instruct, alpha=0.82026.04 | 20.1 | 20.4 | 63.4 | 57.9 | |
| steeringModel=Llama-3.2-3B-Instruct, alpha=0.42026.04 | 23.3 | 24.9 | 54.1 | 45 | |
| mean ablationModel=Llama-3.2-3B-Instruct2026.04 | 24.5 | 27.3 | 51.7 | 39.8 | |
| steeringModel=Qwen2.5-3B-Instruct, alpha=0.32026.04 | 29.5 | 24.4 | 46.5 | 49.6 | |
| steeringModel=Llama-3.2-3B-Instruct, alpha=0.32026.04 | 31.8 | 29.9 | 37.4 | 34.1 | |
| mean ablationModel=Qwen2.5-3B-Instruct2026.04 | 37.4 | 30.4 | 32.1 | 37.3 | |
| baselineModel=Llama-3.2-3B-Instruct2026.04 | 50.7 | 45.3 | — | — | |
| baselineModel=Qwen2.5-3B-Instruct2026.04 | 55.1 | 48.5 | — | — |