Convex Compositional Reasoning Models
The paper introduces Convex Compositional Energy Minimization (CCEM) to solve complex reasoning problems using energy-based models. It overcomes non-convexity by parameterizing local factors with input-convex neural networks, ensuring that the global objective remains convex under summation. This allows for deterministic first-order optimization instead of sampling, enabling reliable inference and strong generalization to larger combinatorial problems.