MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention

arxiv:2506.13585

Original source · Paper JSON

Post-training · 2025-06-16

Cites (5) · Cited by (12)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

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

Contents

• MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention
• Abstract
• 1 Introduction
• 2 Preparation for Scalable RL: Continual Pretraining and SFT
• 2.1 Continual Pre-Training: Foundation for RL Scaling
• 2.2 Supervised Fine-Tuning: Focused Alignment for Efficient RL
• 3 Efficient RL Scaling: Algorithms and Lightning Attention
• 3.1 Efficient RL Scaling with CISPO
• 3.2 Efficient RL Scaling with Lightning Attention – Challenges and Recipes
• 4 Scaling Reinforcement Learning with Diverse Data
• 4.1 Reasoning-Intensive Tasks with Rule-based Verification
• 4.2 General Domain Tasks with Model-based Feedbacks
• 4.2.1 Data and Reward Models
• 4.2.2 Addressing Bias of Generative Reward Models for Long CoT
• 4.3 Curriculum of Incorporating Diverse Data
• 5 Extending RL Scaling to Longer Thinking
• 6 Evaluations
• 6.1 Core Benchmarks
• 6.2 Effect of RL Scaling
• 7 Conclusion and Future work
• Appendix A Contributors
• Instructions for reporting errors