LLM Zeroth-Order Fine-Tuning Is an Inference Workload, Not a Training One

arXiv·medium signal

Shows that zeroth-order fine-tuning of LLMs — which replaces backpropagation with forward-pass objective evaluations — should be reimplemented as an inference workload rather than running inside conventional training loops. This reframing unlocks existing inference infrastructure (batching, KV caching, quantization) for fine-tuning. Practical cost reduction for teams fine-tuning without gradient access.