Learning to Discover at Test Time

arxiv:2601.16175

Original source · Paper JSON

Post-training · 2026-01-22

Cites (4) · Cited by (6)

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

• Learning to Discover at Test Time
• Abstract
• 1 Introduction
• 2 Preliminaries
• 2.1 Discovery Problem
• Definition (Discovery) .
• 2.2 Search Methods
• 3 Learning to Discover at Test Time
• 3.1 Naive RL at Test Time
• 3.2 TTT-Discover
• 3.3 Implementation Details
• 4 Applications
• 4.1 Mathematics
• 4.1.1 Erdős’ Minimum Overlap Problem
• 4.1.2 Autocorrelation Inequalities
• 4.1.3 Circle Packing
• 4.1.4 Expert Review
• 4.2 Kernel Engineering
• 4.2.1 Expert Review
• 4.3 Algorithm Engineering
• 4.4 Single Cell Analysis
• 4.4.1 Expert Review
• 4.5 Ablations
• 5 Related Works
• 5.1 Continual Learning
• 5.2 Test-Time Training
• 5.2.1 TTT on Nearest Neighbors: Larger Effective Capacity
• 5.2.2 TTT for Novel Instances: Better Generalization
• 5.3 RL on One Example
• 5.4 RL on the Test Set
• 6 Future Work
• Acknowledgments
• Appendix A Training details
• A.1 Entropic utility objective
• Discussion.
• A.2 PUCT Prioritization
• Comparison to AlphaZero PUCT.
• Appendix B Mathematics
• B.1 Circle Packing
• B.2 Autocorrelation Inequalities
• B.3 Erdős’
• Appendix C Kernel engineering
• C.1 Kernel evaluation details
• C.2 Analysis of best generated kernels
• C.3 TTT MLA-Decode kernels filtered with Triton kernels
• Appendix D Algorithm Engineering
• Appendix E Single cell analysis
• Appendix F Prompts
• Instructions for reporting errors