Compositional Steering on 2-behavior combinations (unseen split)
76.9AccuracyComposition (<and>)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Composition (<and>)Backbone=Qwen-8B2026.01 | 76.9 | 5.3 | 4.9 | |
| Hybrid: Composition (<and>) + Instr.Backbone=Qwen-8B2026.01 | 76.3 | 4.4 | 4.9 | |
| Instruction SteeringBackbone=Qwen-8B2026.01 | 71.8 | 7.8 | 4.9 | |
| ConcatenationBackbone=Qwen-8B2026.01 | 62.1 | 25.3 | 4.8 | |
| LoRA DAREBackbone=Qwen-8B2026.01 | 44.8 | — | 4.7 | |
| LM-SteerBackbone=Qwen-8B2026.01 | 13.4 | — | 1.3 |