Code Debugging on InfiniteBench code_debug 40k input cap
34.26AccuracyQwen3-4B
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen3-4BBackbone=Qwen3-4B, Mode=Full KV, Input cap=40k2026.05 | 34.26 | |
| NestedKVBackbone=Qwen3-4B, Compression ratio (r)=0.50, Input cap=40k2026.05 | 34.01 | |
| NestedKVBackbone=Qwen3-4B, Compression ratio (r)=0.25, Input cap=40k2026.05 | 33.76 | |
| SnapKVBackbone=Qwen3-4B, Compression ratio (r)=0.25, Input cap=40k2026.05 | 32.49 | |
| StreamingLLMBackbone=Qwen3-4B, Compression ratio (r)=0.25, Input cap=40k2026.05 | 32.49 | |
| KeyDiffBackbone=Qwen3-4B, Compression ratio (r)=0.50, Input cap=40k2026.05 | 32.23 | |
| ExpAttnBackbone=Qwen3-4B, Compression ratio (r)=0.25, Input cap=40k2026.05 | 31.98 | |
| StreamingLLMBackbone=Qwen3-4B, Compression ratio (r)=0.50, Input cap=40k2026.05 | 31.98 | |
| KeyDiffBackbone=Qwen3-4B, Compression ratio (r)=0.25, Input cap=40k2026.05 | 31.47 | |
| NestedKVBackbone=Qwen3-4B, Compression ratio (r)=0.75, Input cap=40k2026.05 | 31.22 | |
| KeyDiffBackbone=Qwen3-4B, Compression ratio (r)=0.75, Input cap=40k2026.05 | 30.96 | |
| ExpAttnBackbone=Qwen3-4B, Compression ratio (r)=0.50, Input cap=40k2026.05 | 30.46 | |
| SnapKVBackbone=Qwen3-4B, Compression ratio (r)=0.50, Input cap=40k2026.05 | 30.2 | |
| StreamingLLMBackbone=Qwen3-4B, Compression ratio (r)=0.75, Input cap=40k2026.05 | 29.44 | |
| SnapKVBackbone=Qwen3-4B, Compression ratio (r)=0.75, Input cap=40k2026.05 | 28.17 | |
| ExpAttnBackbone=Qwen3-4B, Compression ratio (r)=0.75, Input cap=40k2026.05 | 28.17 | |
| PyramidKVBackbone=Qwen3-4B, Compression ratio (r)=0.75, Input cap=40k2026.05 | 26.65 | |
| PyramidKVBackbone=Qwen3-4B, Compression ratio (r)=0.25, Input cap=40k2026.05 | 26.4 | |
| PyramidKVBackbone=Qwen3-4B, Compression ratio (r)=0.50, Input cap=40k2026.05 | 23.86 |