AI Agent Cracks Hex Mesh Puzzle

🔥 What happened Researchers at Berkeley trained an AI agent to decompose 3D solids into hexahedral blocks – a decades-old mesh generation problem. The agent makes over 100 decisions per decomposition and solves 94.4% of held-out test cases. 💡 Why it matters Classical algorithms often choke on combinatorial explosion. The kicker: expert iteration beats reinforcement learning – PPO, GRPO, and VinePPO actually degrade the cloned policy. This could speed up finite element simulations, where high-quality hex meshes are pure gold. ⚡ Our take Finally, a paper that tackles a real engineering problem instead of chasing benchmarks. Ditching RL hype for expert iteration is a quiet jab at training orthodoxy.