Self-play fine-tuning has emerged as a promising approach to improve Large Language Models (LLMs) without additional human annotations. However, existing methods struggle with complex generation tasks requiring long context understanding, where models produce partially correct outputs interleaved with errors. Traditional approaches train on entire sequences uniformly, failing to distinguish between well-predicted and erroneous regions, leading to diluted learning signals and slow convergence. We propose DRIFT (Difference-aware Reinforcement through Iterative Fine-Tuning), a novel self-play framework that selectively trains on prediction differences. DRIFT introduces two key innovations: (1) Difference-Aware Masking (DAM) that identifies and masks common subsequences between model outputs and ground truth, focusing training exclusively on error regions; (2) Occurrence-Aware Loss (OAL) that provides position-invariant vocabulary supervision, complementing the position-sensitive adversarial loss. This dual mechanism enables models to correct both positional and lexical errors effectively. Theoretically, we prove that DRIFT converges when masked distributions align. Empirically, we evaluate DRIFT on diverse summarization benchmarks using Qwen2.5-3B and LLaMA-3.1-8B models. Results show that DRIFT significantly outperforms both supervised fine-tuning (SFT) and self-play fine-tuning (SPIN), achieving up to 16\% improvement on SAMSum dialogue summarization tasks while maintaining general capabilities. Notably, DRIFT breaks the performance ceiling of continued SFT and demonstrates superior efficiency compared to holistic self-play methods, validating that targeted optimization on prediction differences is crucial for structured text generation tasks.

Remote expert assistance is critical for complex quality inspection in industrial manufacturing and emergency maintenance. Current Cross-Reality (XR) solutions face significant limitations: predominantly single-user focused, lack effective AI-human expertise integration, and insufficient evaluation of human factors in collaborative contexts. We developed a Cross-Reality collaborative system integrating AI-assisted defect detection with virtual replicas, enabling real-time collaboration between remote experts and local workers via head-mounted displays. Our controlled evaluation with 26 participants inspecting industrial products demonstrated that our system outperformed traditional and screen-based remote assistance methods, reducing error rates from 20.19% (no assistance) to 2.88% (XR-assisted). The system substantially improved social presence, collaboration efficiency, and usability, with 88.5% of remote experts preferring our solution. Our study advances Industry 5.0 principles by combining human-centric design with intelligent technological integration in industrial quality inspection scenarios.

We present a novel Mixed Reality (MR) remote collaboration system that integrates a telepresence drone to overcome viewpoint limitations in previous approaches. Current remote collaboration systems often restrict remote users to the local user's perspective or a fixed viewpoint, which may reduce spatial awareness and collaboration effectiveness. Our proposed system enables a remote user wearing a VR headset to independently navigate the local environment through a drone while guiding a local AR headset user. The drone represents the remote user physically and virtually with an AR avatar. The system provides the remote user with a 3D reconstruction of the local environment, the local user's avatar, and real-time drone camera video to facilitate drone operation and collaboration. A user study comparing our drone-enhanced MR system with 2D video and 360-degree video MR systems during inspection tasks revealed that the telepresence drone significantly improved collaboration efficiency and enhanced users' social presence and spatial presence. However, the remote users perceived a higher workload. Our findings demonstrate the benefits of independent spatial navigation through drones in remote collaboration and offer insights for future drone-enhanced MR remote collaboration systems.

Augmented Reality (AR) has been used to enhance the shopping experience. However, existing AR shopping systems mainly focus on the solo user’s experience, while lacking multi-user experience. To address the gap, we propose a novel approach to collaborative shopping in a multi-user AR environment. By integrating the interactive personalized avatars of customers and shared gaze cues between shopping companions, we aim to understand how these technologies can enhance the collaborative shopping experience. We recruited thirty participants to conduct a 2 (personalized avatar: static vs. interactive) times 2 (shared gaze: without vs. with) within-subject repeat user study. The quantitative results from questionnaires showed that both interactive personalized avatars and shared gaze cues had positive effects on participants’ perceptions of enjoyment, usefulness, communication, co-presence, and future use. The combination of two features further enhanced the communication and perceived co-presence between shoppers and was preferred by participants. The qualitative results showed that interactive personalized avatars and shared gaze cues can enhance the shopping experience and promote efficiency which is consistent with quantitative results.

In the traditional data-driven fault diagnosis task, gathering training samples for all possible fault classes poses a significant challenge. There are many target faults that cannot be collected in advance, which potentially limiting the performance of fault diagnosis models. Zero-shot learning has emerged as a viable solution to this problem. However, it often encounters the issue of domain shift. In this article, an attribute-consistent generative adversarial network with feature generation (ACGAN-FG) is proposed for zero-shot fault diagnosis. ACGAN-FG introduces a discriminative classifier and a binary comparator to construct the attribute-consistent losses, which can alleviate the issue that the generated features may deviate from real faults. To generate fault features with greater diversity and enhance the robustness of the proposed model, a cycle rank loss is designed. Besides, this method also introduces feature concatenation to build new training data and testing data. This concatenation can transform the generated features into more discriminative representation for further fault diagnosis. The effectiveness of the proposed method is validated on two cases for fault diagnosis purpose. The results also indicate that the proposed method is outperforms other state-of-art zero-shot fault diagnosis methods.