Generative Recursive Reasoning

arxiv:2605.19376

Original source · Paper JSON

Reasoning and the "physics" of language models · 2026-05-19

Cites (8) · Cited by (2)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4
5. Part 5

Contents

• Generative Recursive Reasoning
• Abstract
• 1 Introduction
• 2 Generative Recursive Reasoning Models
• 2.1 Architecture
• 2.2 Training
• 2.3 Inference-Time Scaling
• 3 Related Work
• 4 Experiments
• 4.1 Challenging Puzzle Tasks
• 4.2 Multi-solution Puzzle Tasks
• 4.3 Exploring GRAM as an Unconditional Generator
• 4.4 Ablation Study
• 5 Conclusions and Limitations
• Acknowledgment
• Broader Impacts
• Appendix A Additional Method Details
• A.1 Adaptive Computation Time
• A.2 Latent Process Reward Model (LPRM).
• A.3 Empirical Validation of the Surrogate Objective
• Appendix B Training and Architecture Details
• B.1 Architecture Details
• Encoder and Decoder for Image Patches.
• B.2 Training Details
• Appendix C Additional Details of Experiment Setup
• C.1 Challenging Puzzle Tasks
• C.1.1 Looped TF on ARC-AGI
• C.2 Multi-solution Puzzle Tasks
• C.2.1 N-Queens Problem
• C.2.2 Graph Coloring Problem
• Appendix D Additional Experiment Results
• D.1 Additional Results on Challenging Puzzle Benchmarks
• GRAM Advances the Recursive-Reasoning Line.
• Sudoku-Extreme Resists Parameter Scaling.
• D.2 Scales with Parallel Sampling on ARC-AGI Challenge
• Interaction between Augmentation and Sampling.
• D.3 Solution Coverage Analysis
• D.4 Additional Generated Image Samples
• D.5 Additional Experiment Results on Unconditional Sudoku Generation
• D.6 Visualizing Latent Recursion Process
• Licenses
• Instructions for reporting errors