Evaluating Context Influence and Input Regurgitation on CNN-DM
140Information Influence (\hat{I})OPT 1.3B
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| OPT 1.3BDecoding λ=1.52024.10 | 140 | 559 | 836 | |
| GPT-Neo 1.3BDecoding λ=1.52024.10 | 130.47 | 637 | 822 | |
| LLaMA 3 8BDecoding λ=1.52024.10 | 98.99 | 429 | 882 | |
| OPT 1.3BDecoding λ=1.02024.10 | 85.23 | 373 | 644 | |
| GPT-Neo 1.3BDecoding λ=1.02024.10 | 77.87 | 338 | 571 | |
| LLaMA 3 8BDecoding λ=1.02024.10 | 64.61 | 285 | 632 | |
| OPT 1.3BDecoding λ=0.52024.10 | 17.5 | 1 | 87 | |
| LLaMA 3 8BDecoding λ=0.52024.10 | 15.97 | 8 | 109 | |
| GPT-Neo 1.3BDecoding λ=0.52024.10 | 15.16 | 5 | 87 |