Uncertainty Quantification for LLM Agents: A Taxonomy, an Evaluation Protocol, and an Empirical Study
本文针对大语言模型作为代理时的不确定性问题，提出了一种分类法、评估协议及实证研究方法，通过Trajectory-Checkpoint Expected Calibration Error (TC-ECE)来更精确地量化多轮对话中的不确定性。
0 citationsRead paper
Selected work
Representative Papers
KANResDiff: Learning Local Residual Diffusion via Kolmogorov-Arnold Network for Ambiguous Medical Image Segmentation
This work addresses the limitations of existing fuzzy medical image segmentation methods, which introduce randomness in a fixed manner and lack progressive semantic modeling. To overcome these issues, the authors propose a stage-aware diffusion framework that leverages Kolmogorov–Arnold networks to learn a local residual diffusion process. The approach employs spline-based time embeddings to enable independent temporal encoding, thereby enhancing semantic distinctions across stages. Furthermore, a learnable-weight residual Schrödinger bridge is introduced to inject deterministic priors, facilitating the construction of locally optimal diffusion trajectories. Evaluated on two public datasets, the model achieves state-of-the-art performance, improving the GED and HM-IoU metrics by 16.8% and 7.7%, respectively, while maintaining competitive results on the MDM metric.
0 citationsRead paper
Optimization of BLE Broadcast Mode in Offline Finding Network
To address high neighbor discovery latency and low success rates in Offline Finding Networks (OFNs) based on Bluetooth Low Energy (BLE), this paper proposes CPBIS—a novel, systematic dual-advertising-interval and interval-ratio optimization mechanism tailored for multi-interval advertisers. CPBIS overcomes the limitations of conventional single-fixed-interval approaches by jointly modeling advertising and scanning timing schedules, integrating probabilistic analysis with an efficient parameter search algorithm. Implemented on the nRF52832 platform, CPBIS achieves co-optimization of discovery latency and success rate. Experimental results demonstrate that CPBIS significantly reduces average discovery latency across diverse scanning modes while improving measured discovery success rate by 27%. Furthermore, end-to-end object-finding responsiveness becomes more stable and reliable, enhancing overall OFN performance.
0 citationsRead paper
HGO-YOLO: Advancing Anomaly Behavior Detection with Hierarchical Features and Lightweight Optimized Detection
To address real-time anomaly behavior detection under hardware-constrained environments, this paper proposes HGO-YOLO—a lightweight and efficient model. It adopts HGNetv2 as the backbone and integrates it with the YOLOv8 framework, introducing a novel Hierarchical-Ghost multi-level feature fusion mechanism and a parameter-sharing lightweight detection head, OptiConvDetect, which significantly enlarges the receptive field while reducing model redundancy. Experimental results show that HGO-YOLO achieves a compact size of 4.6 MB, attains 56 FPS on CPU, and delivers an mAP@0.5 of 87.4% with a recall rate of 81.1%. It reduces computational cost by 51.7% and accelerates inference by 1.7× over the YOLOv8 baseline. The model strikes an optimal trade-off among accuracy, speed, and parameter efficiency, and supports deployment across diverse platforms—including CPUs, Raspberry Pi 4, and NVIDIA GPUs.
0 citationsRead paper
Recent publications
Latest Papers
Uncertainty Quantification for LLM Agents: A Taxonomy, an Evaluation Protocol, and an Empirical Study
本文针对大语言模型作为代理时的不确定性问题，提出了一种分类法、评估协议及实证研究方法，通过Trajectory-Checkpoint Expected Calibration Error (TC-ECE)来更精确地量化多轮对话中的不确定性。
0 citationsRead paper
KANResDiff: Learning Local Residual Diffusion via Kolmogorov-Arnold Network for Ambiguous Medical Image Segmentation
This work addresses the limitations of existing fuzzy medical image segmentation methods, which introduce randomness in a fixed manner and lack progressive semantic modeling. To overcome these issues, the authors propose a stage-aware diffusion framework that leverages Kolmogorov–Arnold networks to learn a local residual diffusion process. The approach employs spline-based time embeddings to enable independent temporal encoding, thereby enhancing semantic distinctions across stages. Furthermore, a learnable-weight residual Schrödinger bridge is introduced to inject deterministic priors, facilitating the construction of locally optimal diffusion trajectories. Evaluated on two public datasets, the model achieves state-of-the-art performance, improving the GED and HM-IoU metrics by 16.8% and 7.7%, respectively, while maintaining competitive results on the MDM metric.
0 citationsRead paper
Optimization of BLE Broadcast Mode in Offline Finding Network
To address high neighbor discovery latency and low success rates in Offline Finding Networks (OFNs) based on Bluetooth Low Energy (BLE), this paper proposes CPBIS—a novel, systematic dual-advertising-interval and interval-ratio optimization mechanism tailored for multi-interval advertisers. CPBIS overcomes the limitations of conventional single-fixed-interval approaches by jointly modeling advertising and scanning timing schedules, integrating probabilistic analysis with an efficient parameter search algorithm. Implemented on the nRF52832 platform, CPBIS achieves co-optimization of discovery latency and success rate. Experimental results demonstrate that CPBIS significantly reduces average discovery latency across diverse scanning modes while improving measured discovery success rate by 27%. Furthermore, end-to-end object-finding responsiveness becomes more stable and reliable, enhancing overall OFN performance.
0 citationsRead paper
HGO-YOLO: Advancing Anomaly Behavior Detection with Hierarchical Features and Lightweight Optimized Detection
To address real-time anomaly behavior detection under hardware-constrained environments, this paper proposes HGO-YOLO—a lightweight and efficient model. It adopts HGNetv2 as the backbone and integrates it with the YOLOv8 framework, introducing a novel Hierarchical-Ghost multi-level feature fusion mechanism and a parameter-sharing lightweight detection head, OptiConvDetect, which significantly enlarges the receptive field while reducing model redundancy. Experimental results show that HGO-YOLO achieves a compact size of 4.6 MB, attains 56 FPS on CPU, and delivers an mAP@0.5 of 87.4% with a recall rate of 81.1%. It reduces computational cost by 51.7% and accelerates inference by 1.7× over the YOLOv8 baseline. The model strikes an optimal trade-off among accuracy, speed, and parameter efficiency, and supports deployment across diverse platforms—including CPUs, Raspberry Pi 4, and NVIDIA GPUs.
0 citationsRead paper