Tag
This paper proposes a unified framework called Efficiency Frontier, which treats large model context management as a deployment optimization problem, jointly modeling task performance, token overhead, and preprocessing reuse. On 5,000 HotpotQA instances, deployment optimization saves 25% of token usage, while memory compression is more than half the cost of full context in high-precision scenarios.