Dynamically allocating compute budget to hard set of problems and evolving the sections with Qwen-35B-A3B gets you near GPT-5.4-xHigh on HLE