Things I should try:
- pretrain a very tiny LM from scratch or Build a Large Language Model
- take a pretrained model, and apply finetuning tricks: Practical Tips for Finetuning LLMs Using LoRA (Low-Rank Adaptation) or Tinker
- Take a separate pretrained model, and post train it to be a reasoning model: From Base to Reasoning Model: A Full Post-Training Pipeline on a Single Node | Pluralis Research or - RLHF Book: Reinforcement Learning from Human Feedback and LLM Post-Training or Build a Reasoning Model (From Scratch) - Sebastian Raschka
- NanoGPT Pro — Multi-Architecture NanoGPT Training & Evaluation Suite: I should try to create this kind of architecture zoo as well.
To-Read
- [2510.15511] Language Models are Injective and Hence Invertible
- [2510.13786] The Art of Scaling Reinforcement Learning Compute for LLMs
- [2507.20783] On The Role of Pretrained Language Models in General-Purpose Text Embeddings: A Survey
- [2601.16175] Learning to Discover at Test Time
- COT: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
- [2512.07805] Group Representational Position Encoding
RL:
- [2303.11366] Reflexion: Language Agents with Verbal Reinforcement Learning
- GRPO: [2501.12948] DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- From Base to Reasoning Model: A Full Post-Training Pipeline on a Single Node | Pluralis Research
- State of RL for reasoning LLMs | A. Weers