Mathematical Reasoning on AIME24 (Accuracy, Sequential Tokens, Total Tokens)
73.3AccuracySC@32
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SC@32Backbone=Qwen3-4B-Instruct-Thinking2026.06 | 73.3 | 13,000 | 221,300 | — | — | |
| ASCBackbone=Qwen3-4B-Instruct-Thinking2026.06 | 73.3 | 127,800 | 127,800 | — | — | |
| RL-guidedBackbone=Qwen3-4B-Instruct-Thinking2026.06 | 73 | 30,800 | 90,100 | — | — | |
| ESCBackbone=Qwen3-4B-Instruct-Thinking2026.06 | 72.7 | 48,200 | 161,500 | — | — | |
| SC@32Backbone=Qwen3-1.7B-Thinking2026.06 | 68.3 | 26,200 | 494,100 | — | — | |
| ASCBackbone=Qwen3-1.7B-Thinking2026.06 | 68.2 | 317,700 | 317,700 | — | — | |
| RL-guidedBackbone=Qwen3-1.7B-Thinking2026.06 | 67.6 | 67,600 | 210,200 | — | — | |
| ESCBackbone=Qwen3-1.7B-Thinking2026.06 | 67.3 | 118,700 | 412,800 | — | — | |
| SNAPKV + HARD-KVTop-p budget=0.60, Config=+ Ours2026.06 | 50.2 | — | — | 88.2 | 646 | |
| SNAPKV + HARD-KVTop-p budget=0.90, Config=+ Ours2026.06 | 49.2 | — | — | 90.3 | 632 | |
| SNAPKVTop-p budget=0.90, Config=Base2026.06 | 43.3 | — | — | 30.7 | 301 | |
| SC@32Backbone=GPT-4.1-nano2026.06 | 37.1 | 5,300 | 92,700 | — | — | |
| ASCBackbone=GPT-4.1-nano2026.06 | 37.1 | 68,600 | 68,600 | — | — | |
| ESCBackbone=GPT-4.1-nano2026.06 | 36.9 | 24,100 | 81,100 | — | — | |
| RL-guidedBackbone=GPT-4.1-nano2026.06 | 36.9 | 26,000 | 55,400 | — | — | |
| VANILLA + HARD-KVTop-p budget=0.60, Config=+ Ours2026.06 | 36.7 | — | — | 77.1 | 728 | |
| VANILLATop-p budget=0.90, Config=Base2026.06 | 35.3 | — | — | 33.2 | 279 | |
| VANILLATop-p budget=0.60, Config=Base2026.06 | 33.9 | — | — | 27.1 | 280 | |
| VANILLA + HARD-KVTop-p budget=0.90, Config=+ Ours2026.06 | 33.5 | — | — | 83.4 | 732 | |
| SNAPKVTop-p budget=0.60, Config=Base2026.06 | 26.7 | — | — | 23.4 | 300 | |
| SC@32Backbone=Qwen3-0.6B-Thinking2026.06 | 22.6 | 24,700 | 495,600 | — | — | |
| ASCBackbone=Qwen3-0.6B-Thinking2026.06 | 22.6 | 437,100 | 437,100 | — | — | |
| RL-guidedBackbone=Qwen3-0.6B-Thinking2026.06 | 22.6 | 115,700 | 356,000 | — | — | |
| ESCBackbone=Qwen3-0.6B-Thinking2026.06 | 22.3 | 138,900 | 480,300 | — | — | |
| Base SystemSystem setting=System Throughput Baseline2026.06 | — | — | — | — | 346 | |
| CUDA GraphSystem setting=+ CUDA Graph2026.06 | — | — | — | — | 477 | |
| Top-k pruningSystem setting=+ Top-k pruning2026.06 | — | — | — | — | 806 |