Retrieval-Augmented Generation on RAG QA Suite (HotpotQA, TriviaQA, 2WikiMQA, MuSiQue, NQ)
54.23HotpotQA ScoreGSA
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GSACompression Ratio=8x, KV-cache Reuse=true, Base Model=Llama3.2-1B2026.04 | 54.23 | 64.81 | 70.21 | 13.98 | 37.1 | 48.07 | |
| H-GSACompression Ratio=16x, KV-cache Reuse=true, Base Model=Llama3.2-1B2026.04 | 50.95 | 62.77 | 68.26 | 12.25 | 37.46 | 46.34 | |
| GSACompression Ratio=16x, KV-cache Reuse=true, Base Model=Llama3.2-1B2026.04 | 44.58 | 59.32 | 61.88 | 9.68 | 32.5 | 41.59 | |
| UniGistCompression Ratio=8x, KV-cache Reuse=true, Base Model=Llama3.2-1B2026.04 | 42.74 | 59.98 | 61.51 | 9.02 | 28.48 | 40.35 | |
| KVLinkCompression Ratio=8x, KV-cache Reuse=true, Base Model=Llama3.2-1B2026.04 | 41.55 | 60 | 63.91 | 9.68 | 30.86 | 41.2 | |
| UniGistCompression Ratio=16x, KV-cache Reuse=true, Base Model=Llama3.2-1B2026.04 | 37.2 | 57.65 | 58.14 | 7.53 | 25.42 | 37.19 | |
| KVLinkCompression Ratio=16x, KV-cache Reuse=true, Base Model=Llama3.2-1B2026.04 | 37.11 | 57.97 | 61.43 | 8.03 | 27.84 | 38.48 |