Reading top conference papers, model metrics keep climbing. But when moving these models toward real‑world inference, latency, cost and robustness often become show‑stoppers. Do you think conference evaluation frameworks should put more weight on deployment‑oriented metrics?