Reasoning on Parity-NL (test)
100AccuracyDoLa
Evaluation results
| Method | Links | |
|---|---|---|
| DoLaBackbone=Qwen3-8B-Instruct, Intervention Method=DoLa2026.01 | 100 | |
| TCRBackbone=Qwen3-8B-Instruct, Intervention Method=TCR2026.01 | 100 | |
| TCR-goldBackbone=Qwen3-8B-Instruct, Intervention Method=TCR-gold2026.01 | 100 | |
| Qwen3-8B-InstructBackbone=Qwen3-8B-Instruct, Intervention Method=None2026.01 | 98.7 | |
| TCR-goldBackbone=LLaMA3-8B-Instruct, Intervention Method=TCR-gold2026.01 | 88 | |
| TCRBackbone=LLaMA3-8B-Instruct, Intervention Method=TCR2026.01 | 82 | |
| TCR-goldBackbone=Qwen2.5-7B-Instruct, Intervention Method=TCR-gold2026.01 | 81.2 | |
| DoLaBackbone=LLaMA3-8B-Instruct, Intervention Method=DoLa2026.01 | 71.6 | |
| LLaMA3-8B-InstructBackbone=LLaMA3-8B-Instruct, Intervention Method=None2026.01 | 70 | |
| TCR-goldBackbone=Phi-3-Instruct, Intervention Method=TCR-gold2026.01 | 65.2 | |
| TCRBackbone=Qwen2.5-7B-Instruct, Intervention Method=TCR2026.01 | 60.4 | |
| DoLaBackbone=Qwen2.5-7B-Instruct, Intervention Method=DoLa2026.01 | 58.1 | |
| TCRBackbone=Phi-3-Instruct, Intervention Method=TCR2026.01 | 55.8 | |
| Phi-3-InstructBackbone=Phi-3-Instruct, Intervention Method=None2026.01 | 51.2 | |
| Qwen2.5-7B-InstructBackbone=Qwen2.5-7B-Instruct, Intervention Method=None2026.01 | 48.3 | |
| DoLaBackbone=Phi-3-Instruct, Intervention Method=DoLa2026.01 | 25.8 |