Mathematical Reasoning on GSM8K (test) (Mean |Δs|, # Steerable)
0.06Mean Absolute Difference (|Δs|)Activation Steering (Qwen3-0.6B)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Activation Steering (Qwen3-0.6B)Model=Qwen3-0.6B2026.05 | 0.06 | 1 | |
| Activation Steering (Qwen3-14B)Model=Qwen3-14B2026.05 | 0.17 | 2 | |
| Activation Steering (Qwen3-30B)Model=Qwen3-30B2026.05 | 0.18 | 2 | |
| Activation Steering (Llama-4)Model=Llama-42026.05 | 0.31 | 2 | |
| Activation Steering (Qwen3-235B)Model=Qwen3-235B2026.05 | 0.47 | 4 |