Paper List
- Midtraining Bridges Pretraining and Posttraining Distributions (ICML-2026), https://arxiv.org/pdf/2510.14865
- 从分布角度解释 pretrain, midtrain, posttrain
- PRISM: Demystifying Retention and Interaction in Mid-Training, https://arxiv.org/pdf/2603.17074
- MidTrain 激活模型后续 RL 时的训练表现
- OctoThinker: Mid-training Incentivizes Reinforcement Learning Scaling: https://arxiv.org/pdf/2506.20512
- daVinci-Dev: Agent-native Mid-training for Software Engineering: https://arxiv.org/pdf/2601.18418
Tech Report List
MidTraining -> Continual Pretraining.
- Olmo (13B): https://arxiv.org/pdf/2501.00656
- SmolLM2 (1.7B): https://arxiv.org/pdf/2502.02737
- Qwen3(235B): https://arxiv.org/pdf/2505.09388
- Code World Model (32B): An Open-Weights LLM for Research on Code Generation with World Models, https://arxiv.org/abs/2510.02387
- Qwen3-Coder-Next(80B): https://arxiv.org/pdf/2603.00729
MidTraining Survey: https://arxiv.org/pdf/2510.06826