Mathematical Reasoning on AIME 25 (Accuracy, Throughput, Speedup)
40.8AccuracyFull Attn.
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Full Attn.KV Budget=–, Backbone=Qwen3-8B2026.04 | 40.8 | 222.8 | — | |
| TriAttn.KV Budget=3072, Backbone=Qwen3-8B2026.04 | 40.8 | 563.5 | 2.5 |
| Method | Links | |||
|---|---|---|---|---|
| Full Attn.KV Budget=–, Backbone=Qwen3-8B2026.04 | 40.8 | 222.8 | — | |
| TriAttn.KV Budget=3072, Backbone=Qwen3-8B2026.04 | 40.8 | 563.5 | 2.5 |