Xiaosen Zheng

2 Nov 2026
When do an LLM's RL-usable parameters form?
Across SmolLM3-3B's pretraining, the parameters that reinforcement learning actually uses are laid down early, long before the model can do the task.