The effective detection and governance of Large Language Model (LLM) generated content has become increasingly critical due to the growing risk of misuse. Despite the impressive performance of existing detectors, their reliability and potential in multilingual, real-world scenarios remain largely underexplored.In this study, we introduce DetectRL-X, a comprehensive multilingual benchmark designed to evaluate advanced detectors across 8 dimensions. The benchmark encompasses 8 languages commonly used in commercial contexts and collects human-written texts from 6 domains highly susceptible to LLM misuse. To better aligned with real-world applications, We create LLM-generated texts using 4 popular commercial LLMs, and include typical AI-assisted writing operations such as polishing, expanding, and condensing to capture authentic usage patterns. Furthermore, we develop a multilingual framework for paraphrasing and perturbation attacks to simulate diverse human modifications and writing noise, enabling stress testing of detectors across languages.Experimental results on DetectRL-X reveal the strengths and limitations of current state-of-the-art detectors when applied to diverse linguistic resources. We further analyze how domains, generators, attack strategies, text length, and refinement operations influence performance in different languages, underscoring DetectRL-X as an effective benchmark for strengthening multilingual and language-specific detectors.

Against the backdrop of China’s carbon peak and carbon neutrality goals, the energy system is undergoing rapid transformation. As a vital component, virtual power plants (VPPs) have become a key pathway for integrating distributed energy resources (DERs) into power systems and electricity markets. However, VPP operations still face challenges, such as fluctuating renewable energy output, uncertainty in demand response capabilities, and strategic behavior. To address this, a bi-level optimization scheduling method for VPPs is proposed via Stackelberg game and Q-learning-based differential evolution (QDE). First, a two-stage Stackelberg dynamic game model is established for VPP operators and the user side. VPP operators optimize electricity prices and controllable unit outputs, while industrial users and electric vehicles adjust their loads according to electricity prices to maximize flexibility. Secondly, a solution method combining Q-learning-based differential evolution with quadratic programming is proposed. The upper layer employs QDE optimization, while the lower layer is solved via quadratic programming. Q-learning adaptively adjusts crossover and mutation rates to enhance search capability; parameters are determined through orthogonal experiments, with multi-factor variance analysis validating its effectiveness. Thirdly, sensitivity analysis reveals the impact of carbon trading interval lengths on emissions, identifies the optimal benchmark electricity price, and demonstrates the model’s effectiveness under renewable generation uncertainty. Finally, simulation results demonstrate that this approach reduces carbon emissions and EV charging costs by 20.31% and 8.67%, respectively.

Bo Zeng, Chenyang Lyu, Sinuo Liu, Mingyan Zeng, Minghao Wu, Xuanfan Ni, Tianqi Shi, Yu Zhao, Yefeng Liu, Chenyu Zhu, Ruizhe Li, Jiahui Geng, Qing Li, Yu Tong, Longyue Wang, Weihua Luo, Kaifu Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Large Language Models (LLMs) have demonstrated exceptional performance across diverse natural language processing tasks.However, these models exhibit a critical limitation in output diversity, often generating highly similar content across multiple attempts.This limitation significantly affects tasks requiring diverse outputs, from creative writing to reasoning.Existing solutions, like temperature scaling, enhance diversity by modifying probability distributions but compromise output quality.We propose Guide-to-Generation (G2), a trainingfree plug-and-play method that enhances output diversity while preserving generation quality.G2 employs a base generator alongside dual Guides, which guide the generation process through decoding-based interventions to encourage more diverse outputs conditioned on the original query.Comprehensive experiments demonstrate that G2 effectively improves output diversity while maintaining an optimal balance between diversity and quality.

Despite their outstanding performance in numerous applications, large language models (LLMs) remain prone to hallucinations, generating content inconsistent with their pretraining corpora. Currently, almost all contrastive decoding approaches alleviate hallucinations by introducing a model susceptible to hallucinations and appropriately widening the contrastive logits gap between hallucinatory tokens and target tokens. However, although existing contrastive decoding methods mitigate hallucinations, they lack enough confidence in the factual accuracy of the generated content. In this work, we propose Multi-Model Contrastive Decoding (MCD), which integrates a pretrained language model with an evil model and a truthful model for contrastive decoding. Intuitively, a token is assigned a high probability only when deemed potentially hallucinatory by the evil model while being considered factual by the truthful model. This decoding strategy significantly enhances the model’s confidence in its generated responses and reduces potential hallucinations. Furthermore, we introduce a dynamic hallucination detection mechanism that facilitates token-by-token identification of hallucinations during generation and a tree-based revision mechanism to diminish hallucinations further. Extensive experimental evaluations demonstrate that our MCD strategy effectively reduces hallucinations in LLMs and outperforms state-of-the-art methods across various benchmarks.

Due to the non-renewable character, magnetic materials (especially rare earth permanent magnet) become a hot topic which world scientists are eager to explore and research. Magnetic enterprises are constantly promoting lean production. The core of lean production is production planning and scheduling, which is also the core of enterprise management. It is essential to effective control of plant inventory, improve production efficiency, reduce production cost and rise product delivery satisfaction rate. All above can be realized by using the fast and effective methods which come from multi-process coordination production planning and scheduling. A research method review of magnetic materials multi-process coordination production planning and scheduling is given on the base of describing magnetic materials production process with rare earth as the raw material. Finally, the paper gives four research fields of magnetic materials multi-process coordination production planning and scheduling.

Presentation authoring is an important activity, but often requires the secondary task of collecting the information and media necessary for both slides and speech. Integration of implicit search and peripheral displays into presentation authoring tools may reduce the effort to satisfy not just active needs the author is aware of, but also latent needs that she is not aware of until she encounters content of perceived value. We develop SidePoint, a peripheral panel that supports presentation authoring by showing concise knowledge items relevant to the slide content. We study SidePoint as a technology probe to examine the benefits and issues associated with peripheral knowledge panels for presentation authoring. Our results show that peripheral knowledge panels have the potential to satisfy both types of needs in ways that transform presentation authoring for the better.

MoboQ is a location-based real-time social question answering service deployed in the field in China. Using MoboQ, people can ask temporal and geo-sensitive questions, such as how long is the line at a popular business right now, and then receive answers that crowdsourced from other users in a timely fashion. To obtain answers for questions, the system analyzes the live stream from public microblogging service Sina Weibo to identify people who are likely to currently be at the place that is associated with a question and sends them the unsolicited question through the microblogging service from which they were identified. MoboQ was deployed in China at the beginning of 2012, until October of the same year, it was used to ask 15,224 questions by 35,214 registered users, and it gathered 29,491 answers; 74.6% of the questions received at least one answer, 28% received a first response within 10 minutes, and 51% of the questions got first answer within 20 minutes. In total, 91% of the questions successfully found at least one answer candidate, and they were sent to 162,954 microblogging service users. We analyze the usage patterns and behaviors of the real-world end-users, discuss the lessons learned, and outline the future directions and possible applications that could be built on top of MoboQ.

Media creation applications cater poorly to one very common usage: Situations in which the users need media that they do not own and for which they are unwilling to pay. Finding and using externally produced media is currently a cumbersome process. Often, users locate the content using a search engine, copy it into their work, cross their fingers, and hope they do not infringe on any copyrights. While the authors have shared hundreds of millions of images with permissive licenses, the license terms are too complicated for other users to follow. In our studies, we found that even the well-intentioned users still fail to respect copyrights in simple image reuse situations. We therefore introduce an Open Media Retrieval (OMR) model to remedy this problem and supplement it with prototypes that access various legal image sources directly within the creative work flow and provide automatic credits to the original authors.

In this paper we present a on-field study for evaluating a crowd sourcing mobile social search application. With the help of the local crowd via social medias, this application assists foreign visitors in Japan by answering their image-based questions at hand in a timely fashion. We ran a controlled field experiment for 6 weeks with 55 participants. We found that the mobile crowd sourcing model demonstrated a reliable performance on response speed and response quantity: half of the requests were answered within 10 minutes, 75% of requests were answered within 30 minutes, and on average every request had 4.2 answers. Especially in the afternoon, evening and night, nearly 88% requests were answered in average approximately 10 minutes, with more than 4 answers per request. In terms of participation motivation, we found the top active crowd workers were more driven by intrinsic motivations rather than any of the extrinsic incentives (gamification incentives and social incentives) we designed.

Travelers in countries that use an unfamiliar script cannot use pocket translators or online translation services to understand menus, maps, signs and other important information, because they are unable to write the text they see. Solutions based on optical character recognition provide very limited performance in real-world situations and for complex scripts such as Chinese and Japanese. In this paper, we propose an alternative image translation solution based on crowdsourcing. A large number of human workers on mobile terminals are used to carry out the tasks of image recognition, translation and quality assurance. Compared to purely technical solutions, this human computation approach is also able to account for context and non-textual cues, and provide higher level information to the end-user. In this paper, we describe a preliminary user study to create a model of end-user requirements.