Code Generation on HumanEval+ Out-of-Domain (test)
81.71AccuracyDynaSteer
Evaluation Results
| Method | Links | |
|---|---|---|
| DynaSteerBackbone=Qwen3-8B2026.06 | 81.71 | |
| DRESSBackbone=Qwen3-8B2026.06 | 80.49 | |
| PlainBackbone=Qwen3-8B2026.06 | 79.88 | |
| ACTBackbone=Qwen3-8B2026.06 | 79.88 | |
| ITIBackbone=Qwen3-8B2026.06 | 79.27 | |
| "Wait"Backbone=Qwen3-8B2026.06 | 78.66 | |
| DynaSteerBackbone=Qwen3-1.7B2026.06 | 67.68 | |
| "Wait"Backbone=Qwen3-1.7B2026.06 | 65.85 | |
| ACTBackbone=Qwen3-1.7B2026.06 | 65.85 | |
| PlainBackbone=Qwen3-1.7B2026.06 | 65.24 | |
| ITIBackbone=Qwen3-1.7B2026.06 | 65.24 | |
| DRESSBackbone=Qwen3-1.7B2026.06 | 65.24 | |
| DynaSteerBackbone=Llama3.2-3B-Instruct2026.06 | 54.27 | |
| ACTBackbone=Llama3.2-3B-Instruct2026.06 | 51.83 | |
| DRESSBackbone=Llama3.2-3B-Instruct2026.06 | 51.22 | |
| ITIBackbone=Llama3.2-3B-Instruct2026.06 | 48.78 | |
| "Wait"Backbone=Llama3.2-3B-Instruct2026.06 | 48.17 | |
| PlainBackbone=Llama3.2-3B-Instruct2026.06 | 47.56 |