Teaching large language models (LLMs) to be faithful in the provided context is crucial for building reliable information-seeking systems. Therefore, we propose a systematic framework, CANOE, to reduce faithfulness hallucinations of LLMs across different downstream tasks without human annotations. Specifically, we first synthesize short-form question-answering (QA) data with four diverse tasks to construct high-quality and easily verifiable training data without human annotation. Also, we propose Dual-GRPO, a rule-based reinforcement learning method that includes three tailored rule-based rewards derived from synthesized short-form QA data, while simultaneously optimizing both short-form and long-form response generation. Notably, Dual-GRPO eliminates the need to manually label preference data to train reward models and avoids over-optimizing short-form generation when relying only on the synthesized short-form QA data. Experimental results show that CANOE greatly improves the faithfulness of LLMs across 11 different tasks, even outperforming the most advanced LLMs, e.g., GPT-4o and OpenAI o1.

Kangyang Luo, Yuzhuo Bai, Shuzheng Si, Cheng Gao, Zhitong Wang, Yingli Shen, Wenhao Li, Zhu Liu, Yufeng Han, Jiayi Wu, Cunliang Kong, Maosong Sun. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Shuzheng Si, Haozhe Zhao, Gang Chen, Cheng Gao, Yuzhuo Bai, Zhitong Wang, Kaikai An, Kangyang Luo, Chen Qian, Fanchao Qi, Baobao Chang, Maosong Sun. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

With the advent of network function virtualization (NFV), many users resort to network service provisioning through virtual network function instances (VNFIs) run on the standard physical server in clouds. Following this trend, NFV markets are emerging, which allow a user to procure VNFIs from cloud service providers (CSPs). In such procurement process, it is a significant challenge to ensure differential privacy and truthfulness while explicitly considering dynamic resource provisioning, location sensitiveness and budget of each VNFI. As such, we design a differentially private and truthful reverse auction with dynamic resource provisioning (PTRA-DRP) to resolve the VNFI procurement (VNFIP) problem. To allow dynamic resource provisioning, PTRA-DRP enables CSPs to submit a set of bids and accept as many as possible, and decides the provisioning VNFIs based on the auction outcomes. To be specific, we first devise a greedy heuristic approach to select the set of the winning bids in a differentially privacy-preserving manner. Next, we design a pricing strategy to compute the charges of CSPs, aiming to guarantee truthfulness. Strict theoretical analysis proves that PTRA-DRP can ensure differential privacy, truthfulness, individual rationality, computational efficiency and approximate social cost minimization. Extensive simulations also demonstrate the effectiveness and efficiency of PTRA-DRP.

Diffuse solar radiation (DSR) is essential for carbon absorption in ecosystems and clean energy. Due to the scarcity of DSR observation stations, obtaining spatially continuous and high-accuracy all-sky DSR is a significant challenge. To achieve high-accuracy DSR estimation with limited observational data, this study developed a physics-guided deep learning (DL) algorithm. The algorithm effectively combines the advantages of the radiative transfer model (RTM) and DL and utilizes Himawari-8 top-of-atmosphere (TOA) reflectance and angular data as inputs to estimate DSR. Independent Baseline Surface Radiation Network (BSRN) and Wuhan University station observation validation results show that the algorithm has a high and robust performance in estimating instantaneous (hourly and daily) DSR, with a Pearson correlation coefficient (R) of 0.88 (0.91 and 0.91), a root-mean-square error (RMSE) of 61.84 (50.66 and 17.2) W/m2, and a mean bias error (MBE) of 0.16 (0.5 and −4.43) W/m2. In addition, compared to five existing DSR products (JiEA, CHSSDR, Deep Space Climate Observatory (DSCOVR)/Earth Polychromatic Imaging Camera (EPIC), ERA5, and CERES-SYN1deg), the algorithm shows the highest consistency (hourly$R=0.84$and daily$R=0.86$) and the smallest biases (hourly MBE =9.22 W/m2 and daily MBE =4.9 W/m2) at China Meteorological Administration (CMA) stations. Furthermore, comparisons with Himawari-8’s cloud cover product and related DSR products confirm the spatial rationality and continuity of the estimated DSR by this algorithm. This study demonstrates the advantages of the physics-guided neural network (PGNN) over traditional DL in enhancing the accuracy and transferability of DSR estimation, highlighting its potential for application in DSR retrievals from other similar satellites.

Cloud mask (CM) data are indispensable for estimating downward surface shortwave radiation (DSSR). Most DSSR products are generated using CM data derived from passive satellite observations through the threshold methods. Some uncertainty exists in these CM data, yet the impact of CM quality on the DSSR estimates has received limited attention. To address this gap, this study proposed a method to quantify and mitigate errors in DSSR estimates resulting from CM data error, using the Himawari-8 (H8) products as a case study. First, machine learning (ML) models were constructed for CM, DSSR, and needed atmospheric parameters for DSSR estimation. Then, high-reliability CM estimates were utilized to update the H8 CM. The missing atmospheric parameters resulting from the CM updates were filled by the constructed models. Subsequently, DSSR data were estimated based on the updated CM. Results show that the updated CM effectively corrects misclassifications in the H8 CM, and the differences are more than 600 Wm−2 between DSSR estimates and H8 DSSR for some pixels. Cloud-aerosol Lidar and infrared pathfinder satellite observations (CALIPSO) CM and in situ DSSR were used as truth references for validation. The improved accuracy of the updated CM compared to H8 CM is mainly observed for snow/ice, with clear-sky and cloudy-sky hit rates (HRs) increasing by 0.1 and 0.3, respectively. Besides, when the H8 CM is consistent with the updated CM, the estimated DSSR exhibits a slightly lower root mean square error (RMSE) compared to the H8 DSSR, with a difference of no more than 3 Wm−2. However, in cases where the two CM data are inconsistent, the reduction in RMSE for the estimated DSSR compared to the H8 DSSR is more significant, exceeding 9 Wm−2.