Share your thoughts, 1 month free Claude Pro on usSee more

Downstream Accuracy on MATH500

79.8Accuracy

Victim-Trace (oracle)

Updated 2mo ago

Evaluation Results

Method	Links
Victim-Trace (oracle) 2026.03		79.8
Synthesized-Trace 2026.03		71.8
No finetuning 2026.03		71.2
Victim-Trace (oracle) 2026.03		64.4
Surrogate-Trace 2026.03		63.2
Answer+Summary 2026.03		63
Answer-only 2026.03		61
Synthesized-Trace 2026.03		50.2
Surrogate-Trace 2026.03		48.8
Answer+Summary 2026.03		45.8
Answer-only 2026.03		44
No finetuning 2026.03		42.1