How do you evaluate LLM systems beyond accuracy metrics?

LLM evaluation extends beyond accuracy to include latency, cost efficiency, robustness, safety, and user satisfaction.

Traditional ML metrics like accuracy are insufficient for LLMs. Evaluation includes multi-dimensional metrics: response relevance, hallucination rate, toxicity, latency, token cost, retrieval quality, and user engagement. Systems often combine automated scoring, human feedback, and LLM-as-judge evaluation to get a holistic view.