Selection Is Not Learning: Choosing Model Futures

Benchmark Model Rank Results
mathematical-reasoning-on-aime-2025EvoHeir (Gemma-4-E4B-it)–Accuracy: 37.15
mathematical-reasoning-on-aime-2026EvoHeir (Gemma-4-E4B-it)–Accuracy: 41.88
mathematical-reasoning-on-math-500EvoHeir (Gemma-4-E4B-it)–Accuracy: 86.87
mathematical-reasoning-on-minerva-mathEvoHeir (Qwen3-8B-Base)–Accuracy: 45.22
mathematical-reasoning-on-olympiadbenchEvoHeir (Gemma-4-E4B-it)–Accuracy: 64.14