Downstream Accuracy on MATH500
79.8AccuracyVictim-Trace (oracle)
Evaluation Results
| Method | Links | |
|---|---|---|
| Victim-Trace (oracle)Surrogate=-, Student Model=Qwen2026.03 | 79.8 | |
| Synthesized-TraceSurrogate=R1-Distill, Student Model=Qwen2026.03 | 71.8 | |
| No finetuningSurrogate=-, Student Model=Qwen2026.03 | 71.2 | |
| Victim-Trace (oracle)Surrogate=-, Student Model=Llama2026.03 | 64.4 | |
| Surrogate-TraceSurrogate=R1-Distill, Student Model=Qwen2026.03 | 63.2 | |
| Answer+SummarySurrogate=-, Student Model=Qwen2026.03 | 63 | |
| Answer-onlySurrogate=-, Student Model=Qwen2026.03 | 61 | |
| Synthesized-TraceSurrogate=R1-Distill, Student Model=Llama2026.03 | 50.2 | |
| Surrogate-TraceSurrogate=R1-Distill, Student Model=Llama2026.03 | 48.8 | |
| Answer+SummarySurrogate=-, Student Model=Llama2026.03 | 45.8 | |
| Answer-onlySurrogate=-, Student Model=Llama2026.03 | 44 | |
| No finetuningSurrogate=-, Student Model=Llama2026.03 | 42.1 |