notesum.ai
Published at November 14StreamAdapter: Efficient Test Time Adaptation from Contextual Streams
cs.CL
cs.AI
Released Date: November 14, 2024
Authors: Dilxat Muhtar1, Yelong Shen2, Yaming Yang2, Xiaodong Liu2, Yadong Lu2, Jianfeng Liu2, Yuefeng Zhan2, Hao Sun2, Weiwei Deng2, Feng Sun2, Xueliang Zhang1, Jianfeng Gao2, Weizhu Chen2, Qi Zhang2
Aff.: 1Nanjing University; 2Microsoft

| Seen Task | Unseen Task | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| BoolQ | CoPA | SST2 | CB | RTE | Avg. | Hellaswag | Winogrande | OBQA | ARC-C | ARC-E | PIQA | Avg. | ||
| Zero-shot | Tiny-Llama-1.1B | 57.03 | 78.00 | 69.61 | 14.29 | 51.99 | 54.18 | 59.01 | 58.88 | 21.80 | 27.56 | 60.31 | 73.34 | 50.15 |
| ICL | 63.39 | 76.00 | 78.21 | 51.79 | 49.10 | 63.70 | 59.46 | 59.83 | 26.20 | 30.61 | 64.81 | 73.78 | 52.45 | |
| LoRA | 74.28 | 78.00 | 86.58 | 80.36 | 69.31 | 77.71 | 59.01 | 56.35 | 24.80 | 27.90 | 57.15 | 72.14 | 49.56 | |
| TempLoRA | 57.61 | 74.00 | 71.33 | 39.29 | 56.68 | 59.78 | 59.42 | 60.46 | 21.60 | 27.82 | 62.12 | 72.09 | 50.59 | |
| H2O | 62.72 | 77.00 | 75.80 | 23.21 | 47.23 | 57.19 | 58.86 | 59.12 | 23.20 | 30.20 | 62.58 | 72.96 | 51.15 | |
| SnapKV | 63.29 | 75.00 | 63.30 | 44.64 | 46.93 | 58.63 | 58.09 | 59.83 | 23.40 | 28.75 | 62.92 | 72.80 | 50.97 | |
| StreamAdapter | 81.77 | 76.00 | 89.56 | 85.71 | 82.67 | 83.14 | 59.45 | 59.91 | 27.00 | 31.06 | 64.93 | 73.29 | 52.61 | |
| Zero-shot | LLaMA-3-8B | 81.11 | 88.00 | 67.09 | 51.79 | 67.87 | 71.17 | 79.15 | 72.93 | 35.60 | 50.09 | 80.35 | 79.60 | 66.29 |
| ICL | 83.49 | 89.00 | 95.07 | 83.93 | 79.06 | 86.11 | 81.46 | 78.45 | 36.40 | 52.43 | 82.79 | 80.58 | 68.69 | |
| LoRA | 89.24 | 84.00 | 96.22 | 96.43 | 88.09 | 90.80 | 80.78 | 69.14 | 35.40 | 51.62 | 80.77 | 78.67 | 66.06 | |
| TempLoRA | 81.77 | 91.00 | 90.60 | 76.79 | 70.40 | 82.11 | 79.62 | 76.56 | 34.20 | 50.85 | 81.06 | 79.27 | 66.93 | |
| H2O | 82.48 | 88.00 | 92.43 | 73.21 | 75.81 | 82.39 | 80.57 | 75.77 | 33.80 | 51.45 | 82.62 | 79.98 | 67.37 | |
| SnapKV | 84.46 | 88.00 | 94.41 | 75.00 | 74.01 | 83.18 | 80.41 | 76.95 | 35.20 | 50.77 | 82.74 | 79.65 | 67.62 | |
| StreamAdapter | 90.15 | 89.00 | 96.72 | 98.21 | 89.67 | 92.75 | 80.87 | 78.64 | 38.80 | 53.03 | 83.53 | 80.59 | 69.24 | |
| Zero-shot | Phi-3-Medium | 88.65 | 92.00 | 90.48 | 73.21 | 80.51 | 84.97 | 79.12 | 76.48 | 37.80 | 54.95 | 80.85 | 80.90 | 68.35 |
| ICL | 88.23 | 94.00 | 95.07 | 83.93 | 77.62 | 87.77 | 80.70 | 79.40 | 44.80 | 62.80 | 86.87 | 82.37 | 72.82 | |
| LoRA | 88.59 | 92.00 | 96.22 | 92.86 | 89.53 | 91.84 | 79.36 | 76.87 | 41.80 | 58.60 | 84.22 | 81.18 | 70.34 | |
| TempLoRA | 89.27 | 92.00 | 93.81 | 71.43 | 78.34 | 84.97 | 79.30 | 76.56 | 44.40 | 60.58 | 86.28 | 80.61 | 71.29 | |
| H2O | 83.46 | 94.00 | 94.95 | 80.46 | 79.78 | 86.53 | 80.61 | 76.11 | 40.80 | 62.12 | 86.41 | 82.07 | 71.35 | |
| SnapKV | 78.99 | 89.00 | 70.41 | 80.36 | 68.95 | 77.54 | 80.21 | 72.53 | 41.30 | 61.71 | 83.89 | 79.87 | 69.92 | |
| StreamAdapter | 90.24 | 92.00 | 95.18 | 92.86 | 89.89 | 92.03 | 82.03 | 77.11 | 44.80 | 64.33 | 86.91 | 82.47 | 72.94 | |