LLM Inference Interview Questions, issue 16, Aug 15, 2026
The Warm-Start Trap
Senior ML Engineer interview at Anthropic, and the interviewer asks:
“You’re training a tool-use agent. Do you SFT on expert tool-call traces before RL, or go straight to RL from the base model?”
Don’t say: “Always SFT first, a warm start gives the model a good initialization and RL just refines it.”
Why initializing your agent with SFT before RL guarantees a flatlined reward curve, and why skipping straight to cold-start RL builds a more resilient tool user.
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.