notesum.ai
Published at October 24DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations
cs.LG
cs.AI
Released Date: October 24, 2024
Authors: Aryo Pradipta Gema1, Chen Jin, Ahmed Abdulaal2, Tom Diethe2, Philip Teare1, Beatrice Alex1, Pasquale Minervini1, Amrutha Saseendran3
Aff.: 1University of Edinburgh, United Kingdom; 2University College London, United Kingdom; 3AstraZeneca, United Kingdom

| Model | XSum | MemoTrap | IFEval | NQ-Open | NQ-Swap | ||||
|---|---|---|---|---|---|---|---|---|---|
| ROUGE-L | BERTScore-F1 | factKB | Macro Acc | Micro Acc | Prompt Acc | Instruct Acc | EM | EM | |
| Llama3-8b-Instruct | 19.90 | 67.23 | 47.61 | 65.86 | 64.40 | 70.24 | 78.30 | 69.68 | 60.62 |
| + ITI (Li et al., 2024b) | 13.25 | 59.96 | 34.35 | 62.65 | 58.96 | 52.31 | 63.19 | 56.16 | 51.08 |
| + CAD (Shi et al., 2024) | 18.82 | 67.20 | 67.16 | - | - | - | - | 69.83 | 74.21 |
| + DoLA (low) (Chuang et al., 2023) | 19.82 | 67.19 | 47.21 | 65.27 | 63.69 | 69.69 | 78.18 | 69.68 | 60.77 |
| + DoLA (high) (Chuang et al., 2023) | 19.92 | 67.34 | 48.49 | 64.85 | 63.17 | 70.24 | 78.66 | 69.49 | 60.98 |
| + AD (Chen et al., 2024) | 19.79 | 67.31 | 48.49 | 65.38 | 64.28 | 67.65 | 76.26 | 68.93 | 60.51 |
| + DeCoRe | 19.87 | 67.83 | 64.07 | 69.53 | 69.20 | 69.13 | 78.06 | 70.62 | 64.43 |
| + DeCoRe | 19.45 | 67.69 | 66.10 | 74.14 | 74.87 | 68.39 | 76.38 | 70.66 | 66.08 |
| Llama3-70b-Instruct | 22.41 | 69.77 | 61.32 | 68.47 | 66.52 | 77.45 | 84.41 | 71.07 | 76.11 |
| + ITI (Li et al., 2024b) | 21.64 | 69.46 | 61.33 | 71.24 | 68.73 | 76.71 | 83.69 | 71.90 | 74.76 |
| + CD (Li et al., 2023) | 22.71 | 69.99 | 54.73 | 69.27 | 67.55 | 71.72 | 79.74 | 65.80 | 68.37 |
| + CAD (Shi et al., 2024) | 21.45 | 69.28 | 65.61 | - | - | - | - | 71.83 | 84.70 |
| + DoLA (low) (Chuang et al., 2023) | 22.46 | 69.80 | 61.11 | 67.99 | 65.93 | 77.08 | 84.29 | 71.07 | 75.98 |
| + DoLA (high) (Chuang et al., 2023) | 22.43 | 69.93 | 59.99 | 67.92 | 65.81 | 78.00 | 84.65 | 70.40 | 75.26 |
| + AD (Chen et al., 2024) | 22.49 | 69.91 | 60.57 | 67.51 | 66.44 | 76.89 | 84.41 | 71.15 | 74.02 |
| + DeCoRe | 21.94 | 69.35 | 64.88 | 71.96 | 71.41 | 78.56 | 84.89 | 72.51 | 79.06 |
| + DeCoRe | 21.93 | 69.40 | 65.49 | 74.07 | 73.65 | 78.56 | 84.89 | 72.66 | 79.79 |
| + DeCoRe | 22.28 | 69.34 | 59.57 | 72.11 | 70.58 | 61.37 | 71.46 | 71.26 | 75.90 |