Predicting Answer Correctness on HotpotQA (val)
0.8035AUROCattention divergence
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| attention divergenceModel=Llama-3.2-3B2026.05 | 0.8035 | — | — | |
| attention divergenceModel=Qwen3-4B2026.05 | 0.7963 | — | — | |
| attention divergenceModel=Mistral-7B2026.05 | 0.7782 | — | — |