List your company

Predicting Answer Correctness on GSM8K (val)

0.945AUROCattention divergence
0.7600880.8080940.85610.904106May 6, 2026

Evaluation results

MethodLinks
attention divergenceModel=Qwen3-4B2026.050.945——
attention divergenceModel=Mistral-7B2026.050.7889——
attention divergenceModel=Llama-3.2-3B2026.050.7672——