Ming Li, Pei Chen, Zhenhao Zhang, Tao Yang, Xinyang Zhang, Han Li, Tianyu Cao, Ming Zeng, Zhuofeng Wu, Meng Jiang, Huasheng Li, Lihong Li, Bing Yin. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Learning from preference feedback is essential for aligning large language models (LLMs) with human values and improving the quality of generated responses. However, existing preference learning methods rely heavily on curated data from humans or advanced LLMs, which is costly and difficult to scale. In this work, we present PUGC, a novel framework that leverages implicit human Preferences in unlabeled User-Generated Content (UGC) to generate preference data. Although UGC is not explicitly created to guide LLMs in generating human-preferred responses, it often reflects valuable insights and implicit preferences from its creators that has the potential to address readers’ questions. PUGC transforms UGC into user queries and generates responses from the policy model. The UGC is then leveraged as a reference text for response scoring, aligning the model with these implicit preferences. This approach improves the quality of preference data while enabling scalable, domain-specific alignment. Experimental results on Alpaca Eval 2 show that models trained with DPO and PUGC achieve a 9.37% performance improvement over traditional methods, setting a 35.93% state-of-the-art length-controlled win rate using Mistral-7B-Instruct. Further studies highlight gains in reward quality, domain-specific alignment effectiveness, robustness against UGC quality, and theory of mind capabilities. Our code and dataset are available at https://zhaoxuan.info/PUGC.github.io/.

Large Language Models (LLMs) are increasingly employed in multi-turn conversational tasks, yet their pre-training data predominantly consists of continuous prose, creating a potential mismatch between required capabilities and training paradigms. We introduce a novel approach to address this discrepancy by synthesizing conversational data from existing text corpora. We present a pipeline that transforms a cluster of multiple related documents into an extended multi-turn, multi-topic information-seeking dialogue. Applying our pipeline to Wikipedia articles, we curate DocTalk, a multi-turn pre-training dialogue corpus consisting of over 730k long conversations. We hypothesize that exposure to such synthesized conversational structures during pre-training can enhance the fundamental multi-turn capabilities of LLMs, such as context memory and understanding. Empirically, we show that incorporating DocTalk during pre-training results in up to 40% gain in context memory and understanding, without compromising base performance. DocTalk is available at https://huggingface.co/datasets/AmazonScience/DocTalk.

Fenglin Liu, Zheng Li, Hongjian Zhou, Qingyu Yin, Jingfeng Yang, Xianfeng Tang, Chen Luo, Ming Zeng, Haoming Jiang, Yifan Gao, Priyanka Nigam, Sreyashi Nag, Bing Yin, Yining Hua, Xuan Zhou, Omid Rohanian, Anshul Thakur, Lei Clifton, David A. Clifton. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Smoke detection has been a research focus due to its application value in fire and toxic gas leakage alarms. Here we formulate a novel research paradigm for in-depth smoke analysis: modeling the smoke plume in an image with two semantic keypoints, i.e., the start-point (where the smoke comes from) and the end-point (where the smoke spreads), and localizing the keypoints through a heatmap-based detection method. A specialized dataset is developed for smoke keypoint detection, collecting images online and manually annotating the keypoints. Based on the dataset, we propose a Transformer-based model called SmokePose that employs a hierarchical Transformer encoder and a pure Transformer decoder to detect smoke keypoints in an end-to-end manner. We demonstrated the performance of the proposed SmokePose with comparative experiments and ablation studies. A further discussion on the visualization of the attention maps helps to understand the mechanism of SmokePose and to reveal essential image clues for smoke keypoint detection.

The MIS2-TrueFact is geared towards bringing academic, industry, and government researchers and practitioners together to tackle the challenges in misinformation, misbehavior, and data quality issues on the web with heterogeneous and multi-modal sources of information including texts, images, videos, relational data, social networks, and knowledge graphs.

The Satellite communication network is not limited by geographical factors and is an indispensable part of global interconnection. This paper analyzes a hybrid of cloud computing and edge computing LEO satellite network architecture, where the three-tier network is composed of ground users, LEO satellites, and remote cloud servers. Users can not only forward computing tasks to remote cloud servers through LEO satellites for processing, but also directly offload computing tasks to LEO satellites for processing. Based on this, we study the problem of user computing offloading in LEO satellite network and construct a joint optimization problem of offloading decision and computing resource allocation, which aims to reduce the user processing delay and energy consumption when the total computing resources of edge nodes are limited. This problem is a mixed-integer nonlinear problem, which is difficult to be solved in finite time. With the increase in the number of users, the complexity of the solution is very high. Therefore, we reconstruct the problem based on the linear reconstruction technique and relax the binary variables to transform the original non-convex problem into a convex problem. The simulation results show that the algorithm can effectively reduce user delay and energy consumption compared to the case when the tasks are processed locally.

With the explosively increasing video traffic, ensuring the smooth playback of a video has been a challenging problem especially in the fifth generation (5G) mobile communication system. To improve the quality of experience (QoE) of a video playback, the real-time prediction of the video stuck can be a help. In this paper, we firstly select eight features from different layers to reflect the quality of video playback. Then, two models, long and short term memory (LSTM)-based Prediction Model and Gated recurrent unit(GRU)-based Prediction Model, are proposed to predict the stuck state of playback. Finally, to evaluate the effectiveness of the two proposed prediction models, we present the simulation results of accuracy and loss of the two models. Besides, comparison between traditional methods and the proposed one are provided with performance gain in terms of the accuracy, recall, confusion matrix as well as F1-score.

Classification of slow cortical potential (SCP) signals is crucial for brain-computer interface (BCI) systems. This letter presents a new scheme to improve the classification performance of SCP signals. It consists of two parts: first, by fitting the wavelet coefficients of SCP signals with a second-order polynomial, the SCP trends are extracted; and second, a voting system based on the optimal training parameters of the support vector machines is developed to enhance the classification accuracy (CA). Experimental results reveal that the proposed scheme outperforms the state-of-the-art methods. The CA improvements for the dataset Ia of the BCI competition II and the TJU dataset (the dataset was collected in Tianjin University, termed TJU dataset) are reported.