Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

arxiv:2607.27372

Original source · Paper JSON

Data, training, optimization · 2026-07-29

Cites (3) · Cited by (0)

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
6. Part 6
7. Part 7
8. Part 8

Contents

• Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation
• Abstract
• 1 Introduction
• 2 Background
• 2.1 Mode Forcing
• 2.2 End-to-End Generation
• 2.3 Existing Generative Models
• 3 Explorative Modeling Approach
• 3.1 Explorative Modeling Intuition
• Scalable Training of Latent Variable Models
• A Scalable Way to Resolve Coupling
• Generative Modeling via Search
• Spreading Predictions Across Modes
• Minimizing an Implicit Energy
• 3.2 Forward and Reverse Explorative Modeling
• Forward XM.
• Reverse XM.
• What Forward and Reverse XM Optimize.
• 4 Experimentation and Results
• 4.1 Explorative Modeling as a New Scaling Axis
• Does Exploration Improve Existing Generative Models’ Performance?
• Does Performance Across Modalities Scale with Exploration?
• Does Exploration Improve Generalization?
• Does Exploration Improve State-of-the-Art Recipes at Scale?
• Do Improvements From Exploration Vary With Scale?
• Which Models Benefit Most From Exploration?
• 4.2 End-to-End Explorative Models
• Can Explorative Modeling be Used for End-to-End Generation?
• 5 Discussion
• Mode Forcing as a Predictive Theory.
• The Benefits of Surplus Generative Expressivity.
• 6 Future Works and Broader Impact
• Exploration as a Scaling Axis for More Generative Models.
• End-to-end XM Applications.
• Improving and Understanding XMs.
• Scaling Reverse XMs.
• Exploration beyond Pretraining.
• 7 Limitations and Conclusion
• Limitations.
• Conclusion.
• Author Contributions
• Alexi Gladstone
• Heng Ji and Yilun Du
• Acknowledgement
• Appendix A Additional Experimentation
• A.1 Exact Results for Exploration Scaling
• A.2 Comparing Exploration to Minibatch Optimal Transport Couplings
• A.3 Reverse XM Language Models
• Appendix B Additional Intuition
• Generative Modeling with a For Loop.
• Exploration as Building an Associative Memory.
• Appendix C Approach Details
• Batching Forward XM.
• Other Ways to Explore.
• Gradients and Memory.
• Exploration FLOP Cost.
• Generating the K K Candidates.
• Implementing Reverse XM.
• Appendix D Experimental Details
• Model Sizes.
• Image Generation.
• RAE Image Generation.
• Video Generation.
• Individual Results in More Detail.
• Behavior Cloning.
• Goal-Conditioned World Modeling.
• Language Modeling.
• Appendix E Related Works
• E.1 End-to-End Generative Modeling
• E.2 Coupling
• E.3 Explorative Modeling Based Methods
• Appendix F Additional Theory
• F.1 Generative Expressivity Details
• Scope of E E .
• Per-Prediction Expressivity.
• F.2 What Forward and Reverse XM Optimize
• Overview.
• Setup.
• Proposition 1 (Forward XM is maximum likelihood).
• Proof sketch.
• Remark (the optimum at fixed σ \sigma ).
• The role of K K .
• The hard min.
• Proposition 2 (Reverse XM targets the reverse KL, but collapses without an entropy term).
• Proof sketch.
• Where the blur sits.
• On the ELBO.
• Assumptions.
• Proposition 3 (Expressivity of smooth Forward XM under separation).
• Proof sketch.
• Appendix G Frequently Asked Questions
• What’s the main takeaway? How is this paper ‘novel’ if it’s just best-of- K K ?
• Does exploration make inference more expensive?
• Is exploration really a new “axis” if it just costs more compute?
• Why the name Explorative Modeling?
• When does exploration help?
• Does training on the model’s own generations cause collapse?
• How should the amount of exploration K K be chosen?
• Is Explorative Modeling a form of reinforcement learning?
• Instructions for reporting errors