Mathematical Reasoning on AIME (Accuracy, TPR, Efficiency)
237.5TPR (s)ThinKV
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ThinKVModel=R1-Llama-8B, Token Budget=5122025.10 | 237.5 | 46.7 | 0.21 | |
| R-KV (ovl)Model=R1-Llama-8B, Token Budget=5122025.10 | 240.8 | 40 | 0.17 | |
| R-KV (seq)Model=R1-Llama-8B, Token Budget=5122025.10 | 242.6 | 40 | 0.17 | |
| ThinKVModel=R1-Llama-8B, Token Budget=10242025.10 | 243.6 | 50 | 0.22 | |
| R-KV (ovl)Model=R1-Llama-8B, Token Budget=10242025.10 | 246 | 46.7 | 0.2 | |
| R-KV (seq)Model=R1-Llama-8B, Token Budget=10242025.10 | 247.8 | 46.7 | 0.2 | |
| ThinKVModel=R1-Llama-8B, Token Budget=20482025.10 | 251 | 50 | 0.21 | |
| R-KV (ovl)Model=R1-Llama-8B, Token Budget=20482025.10 | 253.7 | 50 | 0.2 | |
| R-KV (seq)Model=R1-Llama-8B, Token Budget=20482025.10 | 254.2 | 50 | 0.2 | |
| FullKVModel=R1-Llama-8B, Token Budget=–2025.10 | 259.6 | 50 | 0.2 |