Mathematical Reasoning on AIME-24 (Avg. Pass@1)
73.75Avg. Pass@1Baseline (Dense)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Baseline (Dense)Backbone=Qwen3-8B2025.12 | 73.75 | 14,400 | |
| KascadeBackbone=Qwen3-8B2025.12 | 70.42 | 15,900 | |
| Kascade (All Heads Pooled)Backbone=Qwen3-8B2025.12 | 65.83 | 17,900 | |
| LessIsMoreBackbone=Qwen3-8B2025.12 | 60.83 | 17,900 | |
| Baseline (Dense)Backbone=DeepSeek-R1-Distill-Llama-8B2025.12 | 50.42 | 11,300 | |
| KascadeBackbone=DeepSeek-R1-Distill-Llama-8B2025.12 | 47.92 | 14,600 | |
| Kascade (All Heads Pooled)Backbone=DeepSeek-R1-Distill-Llama-8B2025.12 | 41.25 | 14,000 | |
| OmniKVBackbone=DeepSeek-R1-Distill-Llama-8B2025.12 | 39.58 | 12,500 | |
| LessIsMoreBackbone=DeepSeek-R1-Distill-Llama-8B2025.12 | 36.25 | 14,800 | |
| QuestBackbone=Qwen3-8B2025.12 | 25.33 | 28,800 | |
| QuestBackbone=DeepSeek-R1-Distill-Llama-8B2025.12 | 7.5 | 22,900 | |
| StreamingLLMBackbone=DeepSeek-R1-Distill-Llama-8B2025.12 | 0 | 7,500 | |
| StreamingLLMBackbone=Qwen3-8B2025.12 | 0 | 6,900 |