mlsanityData to 14 Oct 2026

Convex Compositional Reasoning Models

arXiv 2605.23395 ↗First posted 22 May 2026cs.LGcited by 0

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

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.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 19 of its 23 references that are in the map. First posted on arXiv 22 May 2026.

Builds on

19 of 23 references.

References · 19

  1. 2016Semi-Supervised Classification with Graph Convolutional Networkscited by 11,361
  2. 2017Graph Attention Networkscited by 8,923
  3. 2014Neural Turing Machinescited by 644
  4. 2016Neural Combinatorial Optimization with Reinforcement Learningcited by 608
  5. 2016Input Convex Neural Networkscited by 414
  6. 2019An Efficient Graph Convolutional Network Technique for the Travelling Salesman Problemcited by 269
  7. 2023DIFUSCO: Graph-based Diffusion Solvers for Combinatorial Optimizationcited by 221
  8. 2019Optimal transport mapping via input convex neural networkscited by 206
  9. 2018Learning a SAT Solver from Single-Bit Supervisioncited by 200
  10. 2015Neural GPUs Learn Algorithmscited by 121
  11. 2021Can You Learn an Algorithm? Generalizing from Easy to Hard Problems with Recurrent Networkscited by 117
  12. 2022Generative Flow Networks for Discrete Probabilistic Modelingcited by 114
7 more
  1. 2021Do Neural Optimal Transport Solvers Work? A Continuous Wasserstein-2 Benchmarkcited by 96
  2. 2018Optimal Control Via Neural Networks: A Convex Approachcited by 85
  3. 2025Fast T2T: Optimization Consistency Speeds Up Diffusion-Based Training-to-Testing Solving for Combinatorial Optimizationcited by 45
  4. 2019Graph Colouring Meets Deep Learning: Effective Graph Neural Network Models for Combinatorial Problemscited by 40
  5. 2022NSNet: A General Neural Probabilistic Framework for Satisfiability Problemscited by 18
  6. 2020XLVIN: eXecuted Latent Value Iteration Netscited by 17
  7. 2025Generalizable Reasoning through Compositional Energy Minimizationcited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Learning to Reason with Curriculum II: Compositional Generalization
  2. 2026Reasoning in a Combinatorial and Constrained World: Benchmarking LLMs on Natural-Language Combinatorial Optimization
  3. 2026Reasoning as Energy Minimization over Structured Latent Trajectories
  4. 2026ConvexBench: Can LLMs Recognize Convex Functions?
  5. 2026When Is Compositional Reasoning Learnable from Verifiable Rewards?
  6. 2026Compositional Reasoning in Language Models under Reinforcement Learning Post-Training
  7. 2024Do Large Language Models Have Compositional Ability? An Investigation into Limitations and Scalability
  8. 2026Composition Collapse: Stable Factual Knowledge Does Not Imply Compositional Reasoning