Zirui Wang

Bowen Jin, TJ Collins, Donghan Yu et al.

Large language models (LLMs) exhibit complementary strengths across domains and come with varying inference costs, motivating the design of multi-agent LLM systems where specialized models collaborate efficiently. Existing approaches predominantly rely on decentralized frameworks, which invoke multiple LLMs for every input and thus lead to substantial and uncontrolled inference costs. In this work, we introduce a centralized multi-LLM framework, where a controller LLM selectively coordinates a pool of expert models in a cost-efficient and cost-controllable manner. We formulate this coordination problem as reinforcement learning with dual objectives: maximizing task performance while minimizing the overall inference cost. In addition, we expect the multi-agent system to have adapted behavior with different budget conditions during inference. To this end, we propose CoRL, a reinforcement learning framework that optimizes the performance cost trade-off in a controllable multi-budget setting. Experiments on four diverse benchmarks demonstrate that CoRL enables a single system to surpass the best expert LLM under high-budget settings, while maintaining strong performance in more economical low-budget modes, highlighting the effectiveness of centralized coordination for scalable and cost-efficient multi-agent LLM systems.

Hongyu Li, Jihe Wang, Yu Zhang et al.

A reasonable evaluation standard underlies construction of effective deep learning models. However, we find in experiments that the automatic crack detectors based on deep learning are obviously underestimated by the widely used mean Average Precision (mAP) standard. This paper presents a study on the evaluation standard. It is clarified that the random fractal of crack disables the mAP standard, because the strict box matching in mAP calculation is unreasonable for the fractal feature. As a solution, a fractal-available evaluation standard named CovEval is proposed to correct the underestimation in crack detection. In CovEval, a different matching process based on the idea of covering box matching is adopted for this issue. In detail, Cover Area rate (CAr) is designed as a covering overlap, and a multi-match strategy is employed to release the one-to-one matching restriction in mAP. Extended Recall (XR), Extended Precision (XP) and Extended F-score (Fext) are defined for scoring the crack detectors. In experiments using several common frameworks for object detection, models get much higher scores in crack detection according to CovEval, which matches better with the visual performance. Moreover, based on faster R-CNN framework, we present a case study to optimize a crack detector based on CovEval standard. Recall (XR) of our best model achieves an industrial-level at 95.8, which implies that with reasonable standard for evaluation, the methods for object detection are with great potential for automatic industrial inspection.