Question Answering on 2WikiMultihopQA LongBench
59.73F1 ScoreCORE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| COREGenerator LLM=Qwen2.5-14B-Instruct, Compressor model=Qwen2.5-1.5B-Instruct2025.08 | 59.73 | 108 | |
| Full ContextGenerator LLM=Qwen2.5-14B-Instruct2025.08 | 57.97 | 4,887 | |
| LoopGuardKV budget=1024 tokens2026.04 | 53.61 | — | |
| LongLLMLinguaGenerator LLM=Qwen2.5-14B-Instruct, Compressor model=Qwen2.5-1.5B-Instruct2025.08 | 53.26 | 489 | |
| Full CacheKV budget=1024 tokens2026.04 | 50.49 | — | |
| No ContextGenerator LLM=Qwen2.5-14B-Instruct2025.08 | 43.95 | 0 | |
| DOSToken Budget=128×52026.01 | 41.1 | — | |
| M-RAG†-PToken Budget=128×5, Prompting=Few-shot, Sorting Strategy=Position sorting2026.01 | 39.48 | — | |
| PICToken Budget=128×52026.01 | 39.47 | — | |
| Fixed-SizeToken Budget=128×52026.01 | 38.32 | — | |
| SemanticToken Budget=128×52026.01 | 38.31 | — | |
| M-RAG-SToken Budget=128×5, Prompting=Zero-shot, Sorting Strategy=Similarity sorting2026.01 | 35.82 | — | |
| M-RAG†-SToken Budget=128×5, Prompting=Few-shot, Sorting Strategy=Similarity sorting2026.01 | 35.6 | — | |
| M-RAG-PToken Budget=128×5, Prompting=Zero-shot, Sorting Strategy=Position sorting2026.01 | 34.94 | — | |
| H2OKV budget=1024 tokens2026.04 | 33.73 | — | |
| M-RAG†-PToken Budget=128×3, Prompting=Few-shot, Sorting Strategy=Position sorting2026.01 | 33.01 | — | |
| M-RAG†-SToken Budget=128×3, Prompting=Few-shot, Sorting Strategy=Similarity sorting2026.01 | 32.9 | — | |
| DOSToken Budget=128×32026.01 | 32.55 | — | |
| Fixed-SizeToken Budget=128×32026.01 | 32.39 | — | |
| M-RAG-SToken Budget=128×3, Prompting=Zero-shot, Sorting Strategy=Similarity sorting2026.01 | 32.37 | — | |
| PICToken Budget=128×32026.01 | 32.36 | — | |
| M-RAG-PToken Budget=128×3, Prompting=Zero-shot, Sorting Strategy=Position sorting2026.01 | 32.29 | — | |
| SemanticToken Budget=128×32026.01 | 31.25 | — | |
| M-RAG†-SToken Budget=128×1, Prompting=Few-shot, Sorting Strategy=Similarity sorting2026.01 | 23.21 | — | |
| M-RAG†-PToken Budget=128×1, Prompting=Few-shot, Sorting Strategy=Position sorting2026.01 | 22.32 | — | |
| Fixed-SizeToken Budget=128×12026.01 | 22.2 | — | |
| M-RAG-SToken Budget=128×1, Prompting=Zero-shot, Sorting Strategy=Similarity sorting2026.01 | 21.7 | — | |
| M-RAG-PToken Budget=128×1, Prompting=Zero-shot, Sorting Strategy=Position sorting2026.01 | 21.34 | — | |
| StreamingLLMKV budget=1024 tokens2026.04 | 21.05 | — | |
| PICToken Budget=128×12026.01 | 20.66 | — | |
| DOSToken Budget=128×12026.01 | 20.23 | — | |
| SemanticToken Budget=128×12026.01 | 18.75 | — |