Source Attribution on MultiHop-RAG (test)
98.2Comparison AccuracyGPT-3.5
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GPT-3.5Strategy=Direct Prompting2026.07 | 98.2 | 91.8 | 81.8 | 0 | 80.9 | 91.7 | |
| substring matchType=model-free2026.07 | 97 | 90.2 | 82.7 | 66.8 | 88 | 90.8 | |
| Hidden-State ProbeBackbone=SmolLM2-360M, Strategy=Out-of-fold2026.07 | 94.5 | 91.9 | 81.8 | 92 | 90.5 | 90.3 | |
| Hidden-State ProbeBackbone=Qwen2.5-1.5B, Strategy=Out-of-fold2026.07 | 94.4 | 91.1 | 84.2 | 93.7 | 90.9 | 90.6 | |
| Hidden-State ProbeBackbone=Qwen2.5-0.5B, Strategy=Out-of-fold2026.07 | 93.2 | 90.6 | 82.7 | 91.4 | 89.7 | 89.5 | |
| Hidden-State ProbeBackbone=SmolLM2-135M, Strategy=Out-of-fold2026.07 | 92.4 | 90.9 | 82.8 | 89 | 89.4 | 89.4 |