Compositional Steering on 3-behavior combinations (unseen split)
62.9AccuracyHybrid: Composition (<and>) + Instr.
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Hybrid: Composition (<and>) + Instr.Backbone=Qwen-8B2026.01 | 62.9 | 15.2 | 4.9 | |
| Composition (<and>)Backbone=Qwen-8B2026.01 | 59.5 | 25.5 | 4.9 | |
| Instruction SteeringBackbone=Qwen-8B2026.01 | 54 | 18.1 | 4.9 | |
| ConcatenationBackbone=Qwen-8B2026.01 | 33.2 | 55.8 | 4.8 | |
| LoRA DAREBackbone=Qwen-8B2026.01 | 17.6 | — | 4.6 | |
| LM-SteerBackbone=Qwen-8B2026.01 | 2.1 | — | 1.2 |