Needle-in-a-haystack Retrieval on RULER NIAH
100Retrieval Success (1K)Baseline
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| BaselineModel=LLaMA-3.1-8B-Instruct, Steering Task=None (Unsteered), Steering Type=None2026.05 | 100 | 100 | 99.8 | 98.3 | 97.8 | |
| SKOPModel=LLaMA-3.1-8B-Instruct, Steering Task=Formatting steering, Steering Type=Unconditional2026.05 | 99.8 | 99.5 | 95 | 92.1 | 88.2 | |
| DISCO-QModel=LLaMA-3.1-8B-Instruct, Steering Task=Formatting steering, Steering Type=Unconditional2026.05 | 99.5 | 99.5 | 89.1 | 81.7 | 72.6 | |
| Comm SteerModel=LLaMA-3.1-8B-Instruct, Steering Task=Formatting steering, Steering Type=Unconditional2026.05 | 98.3 | 91.2 | 79.9 | 68.2 | 55.7 |