Mathematical Reasoning on MATH500 (Acc., #Tok.)
47AccuracyFull
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| FullModel=R1-Distill-Qwen-14B, Cache Budget=Full2025.12 | 47 | — | |
| SnapKV-DModel=R1-Distill-Qwen-14B, Cache Budget=5122025.12 | 45 | — | |
| SnapKV-DModel=R1-Distill-Qwen-14B, Cache Budget=2562025.12 | 44 | — | |
| SnapKV-DModel=R1-Distill-Qwen-14B, Cache Budget=1282025.12 | 43 | — | |
| SnapKV-DModel=R1-Distill-Qwen-14B, Cache Budget=3842025.12 | 42 | — | |
| ShadowKVModel=R1-Distill-Qwen-14B, Cache Budget=Fixed2025.12 | 38 | — | |
| StreamingLLMModel=R1-Distill-Qwen-14B, Cache Budget=5122025.12 | 35 | — | |
| H2OModel=R1-Distill-Qwen-14B, Cache Budget=3842025.12 | 31 | — | |
| H2OModel=R1-Distill-Qwen-14B, Cache Budget=5122025.12 | 31 | — | |
| H2OModel=R1-Distill-Qwen-14B, Cache Budget=2562025.12 | 27 | — | |
| StreamingLLMModel=R1-Distill-Qwen-14B, Cache Budget=3842025.12 | 26 | — | |
| H2OModel=R1-Distill-Qwen-14B, Cache Budget=1282025.12 | 20 | — | |
| StreamingLLMModel=R1-Distill-Qwen-14B, Cache Budget=2562025.12 | 17 | — | |
| R-KVModel=R1-Distill-Qwen-14B, Cache Budget=5122025.12 | 9 | — | |
| R-KVModel=R1-Distill-Qwen-14B, Cache Budget=3842025.12 | 3 | — | |
| KNormModel=R1-Distill-Qwen-14B, Cache Budget=5122025.12 | 2 | — | |
| StreamingLLMModel=R1-Distill-Qwen-14B, Cache Budget=1282025.12 | 2 | — | |
| BaseBackbone=DeepScaleR-1.5B-Preview2026.02 | 0.901 | 3,093 | |
| +RLOOBackbone=DeepScaleR-1.5B-Preview2026.02 | 0.895 | 2,843 | |
| TLMREBackbone=DeepScaleR-1.5B-Preview2026.02 | 0.893 | 2,324 | |
| ThinkPrune-4kBackbone=DeepScaleR-1.5B-Preview2026.02 | 0.892 | 2,641 | |
| +LPBackbone=DeepScaleR-1.5B-Preview2026.02 | 0.891 | 2,036 | |
| DDCABackbone=DeepScaleR-1.5B-Preview2026.02 | 0.888 | 2,082 | |
| DDCABackbone=DeepSeek-R1-Distill-1.5B2026.02 | 0.862 | 3,248 | |
| +RLOOBackbone=DeepSeek-R1-Distill-1.5B2026.02 | 0.861 | 4,166 | |
| ThinkPrune-4kBackbone=DeepSeek-R1-Distill-1.5B2026.02 | 0.855 | 3,235 | |
| TLMREBackbone=DeepSeek-R1-Distill-1.5B2026.02 | 0.855 | 4,209 | |
| BaseBackbone=DeepSeek-R1-Distill-1.5B2026.02 | 0.85 | 4,662 | |
| +LPBackbone=DeepSeek-R1-Distill-1.5B2026.02 | 0.843 | 3,069 | |
| KNormModel=R1-Distill-Qwen-14B, Cache Budget=1282025.12 | 0 | — | |
| KNormModel=R1-Distill-Qwen-14B, Cache Budget=2562025.12 | 0 | — | |
| KNormModel=R1-Distill-Qwen-14B, Cache Budget=3842025.12 | 0 | — | |
| R-KVModel=R1-Distill-Qwen-14B, Cache Budget=1282025.12 | 0 | — | |
| R-KVModel=R1-Distill-Qwen-14B, Cache Budget=2562025.12 | 0 | — |