Why Temperature-0 LLM Inference Still Isn’t Deterministic
A practical look at why identical prompts can diverge under temperature=0, how batch-variant kernels create that behavior, and why batch-invariant operators matter for RL and reproducible experiments.