Convex Compositional Reasoning Models
About
Compositional energy-based models can generalize to larger combinatorial reasoning problems by reusing a learned factor energy across many local constraints. In our paper, we show that a key bottleneck in compositional reasoning is not composition itself, but the non-convex geometry of the learned energy landscape. To solve this problem, we introduce Convex Compositional Energy Minimization (CCEM), a framework that parameterizes each factor with an input-convex neural network and optimizes the composed energy over a tight convex relaxation of the feasible set. Because convexity is preserved under summation, the global relaxed objective remains convex, enabling deterministic projected first-order optimization. CCEM is trained in two stages: factor-level contrastive learning to shape local energy basins, followed by end-to-end refinement through an unrolled projected solver. Our experiments show that our models trained on small subproblems or a single problem size transfer to larger instances without retraining.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| 3-SAT Solving | 3-SAT 100 sampled formulas 20 variables, 91 clauses (phase transition 4.258 * n) (test) | Satisfied Clauses Rate96.65 | 11 | |
| 8-Queens Puzzle Solving | 8-Queens (100 sampled boards) | Correct Solutions100 | 7 | |
| Graph Coloring | Erdos Renyi 100 sampled tasks | Average Violations2.2 | 6 | |
| Graph Coloring | Erdos Renyi 2 (100 sampled tasks) | Average Violations4.6 | 6 | |
| Graph Coloring | Holme Kim 100 sampled tasks | Average Violations5 | 6 | |
| Graph Coloring | Holme Kim 2 (100 sampled tasks) | Average Violations14 | 6 | |
| Graph Coloring | Regular Expander 100 sampled tasks | Average Violations4.2 | 6 | |
| Graph Coloring | Paley 100 sampled tasks | Average Violations1.8 | 6 | |
| Graph Coloring | Complete (100 sampled tasks) | Average Violations0.00e+0 | 6 | |
| Graph Coloring | Regular Expander 2 (100 sampled tasks) | -- | 5 |