BDH-CQ: In-Context Learning with Recurrent Latent Reasoning

arxiv:2608.09888

Original source · Paper JSON

Post-training · 2026-08-10

Cites (6) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3

Contents

• BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
• Abstract
• 1 Introduction
• 2 ARC as a controlled, verifiable testbed for BDH-CQ’s in-context generalization
• 3 Introducing BDH-CQ: In-context learning through recurrent memory and latent reasoning
• 3.1 BDH provenance
• 3.2 In-context learning through recurrent memory
• 3.3 Recurrent latent reasoning
• 4 Training data and objective
• 4.1 ARC task formulation
• 4.2 Training mixture
• 5 Public ARC-AGI-1 evaluation
• Independent evaluation.
• 6 Behavioral analysis on ARC-like tasks
• 6.1 Concept-organized capability profile
• 6.2 Simple operators extrapolate; ordering and nesting expose distinct boundaries
• 6.3 Context binds dense mappings, but composition is operation-dependent
• 6.4 Within-task consistency
• 6.5 Identifier and batch-context replication
• 6.6 Attempt delivery, repeatability, and effort
• 7 Scaling reasoning effort increases pass@2
• 8 Related work
• Verbalized reasoning.
• Continuous thoughts in language models.
• Compressed discrete and communicated latent states.
• Recurrent depth and looped models.
• Task-trained recursive solvers.
• ARC and concept-organized evaluation.
• 9 Discussion
• 9.1 Outlook
• 9.2 Scaling model size
• 10 Conclusion
• Appendix A Evaluation set analysis
• A.1 Evaluation sets and surface descriptors
• A.2 Transformation mechanics
• A.3 Controlled structural manipulations
• A.4 Failure structure and scope
• Instructions for reporting errors