List your company

Mathematical Reasoning on AIME24 (Accuracy, Sequential Tokens, Total Tokens)

73.3AccuracySC@32
20.2634.0347.861.57Jun 2, 2026Jun 6, 2026Jun 10, 2026Jun 14, 2026Jun 18, 2026Jun 22, 2026Jun 27, 2026

Evaluation results

MethodLinks
SC@32Backbone=Qwen3-4B-Instruct-Thinking2026.0673.313,000221,300——
ASCBackbone=Qwen3-4B-Instruct-Thinking2026.0673.3127,800127,800——
RL-guidedBackbone=Qwen3-4B-Instruct-Thinking2026.067330,80090,100——
ESCBackbone=Qwen3-4B-Instruct-Thinking2026.0672.748,200161,500——
SC@32Backbone=Qwen3-1.7B-Thinking2026.0668.326,200494,100——
ASCBackbone=Qwen3-1.7B-Thinking2026.0668.2317,700317,700——
RL-guidedBackbone=Qwen3-1.7B-Thinking2026.0667.667,600210,200——
ESCBackbone=Qwen3-1.7B-Thinking2026.0667.3118,700412,800——
SNAPKV + HARD-KVTop-p budget=0.60, Config=+ Ours2026.0650.2——88.2646
SNAPKV + HARD-KVTop-p budget=0.90, Config=+ Ours2026.0649.2——90.3632
SNAPKVTop-p budget=0.90, Config=Base2026.0643.3——30.7301
SC@32Backbone=GPT-4.1-nano2026.0637.15,30092,700——
ASCBackbone=GPT-4.1-nano2026.0637.168,60068,600——
ESCBackbone=GPT-4.1-nano2026.0636.924,10081,100——
RL-guidedBackbone=GPT-4.1-nano2026.0636.926,00055,400——
VANILLA + HARD-KVTop-p budget=0.60, Config=+ Ours2026.0636.7——77.1728
VANILLATop-p budget=0.90, Config=Base2026.0635.3——33.2279
VANILLATop-p budget=0.60, Config=Base2026.0633.9——27.1280
VANILLA + HARD-KVTop-p budget=0.90, Config=+ Ours2026.0633.5——83.4732
SNAPKVTop-p budget=0.60, Config=Base2026.0626.7——23.4300
SC@32Backbone=Qwen3-0.6B-Thinking2026.0622.624,700495,600——
ASCBackbone=Qwen3-0.6B-Thinking2026.0622.6437,100437,100——
RL-guidedBackbone=Qwen3-0.6B-Thinking2026.0622.6115,700356,000——
ESCBackbone=Qwen3-0.6B-Thinking2026.0622.3138,900480,300——
Base SystemSystem setting=System Throughput Baseline2026.06————346
CUDA GraphSystem setting=+ CUDA Graph2026.06————477
Top-k pruningSystem setting=+ Top-k pruning2026.06————806