Mathematical Reasoning on GSM8K (test) (Accuracy, Retained Accuracy, Accuracy Drop)
0AccuracyR-KV (our repro.)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| R-KV (our repro.)Approach=Attn+Redundancy eviction, Offload?=No, Model=R1-Qwen-7B, Context Length=256/512/10242026.05 | 0 | — | — | |
| semantics-aware memory hierarchyApproach=4-tier hierarchy+offload, Offload?=Yes, Model=R1-Qwen-7B, Eviction Ratio=3% evict, n (sample size)=2002026.05 | — | 91 | — | |
| ThinKVApproach=Thought-type quant+evict, Offload?=No, Model=R1-Qwen-14B, Eviction Ratio=5% KV2026.05 | — | — | 4 |