Legal judgment generation is a critical task in legal intelligence. However, existing research in legal judgment generation has predominantly focused on first-instance trials, relying on static fact-to-verdict mappings while neglecting the dialectical nature of appellate (second-instance) review. To address this, we introduce AppellateGen, a benchmark for second-instance legal judgment generation comprising 7,351 case pairs. The task requires models to draft legally binding judgments by reasoning over the initial verdict and evidentiary updates, thereby modeling the causal dependency between trial stages. We further propose a judicial Standard Operating Procedure (SOP)-based Legal Multi-Agent System (SLMAS) to simulate judicial workflows, which decomposes the generation process into discrete stages of issue identification, retrieval, and drafting. Experimental results indicate that while SLMAS improves logical consistency, the complexity of appellate reasoning remains a substantial challenge for current LLMs. The dataset and code are publicly available at: https://anonymous.4open.science/r/AppellateGen-5763.

Emerging technologies such as neural networks, cloud computing, big data, and blockchain have paved the way for the development of artificial intelligence (AI), enabling AI to facilitate business operations. In particular, some organizations seek to leverage AI to replace human agents in positions involving sensitive customer information, with the aim of enhancing privacy protection. However, AI-human interaction tends to fall short of expectations in real-world settings due to the difference between humans and AI. To address this, a study will be conducted to explore the effect of implementing an AI-powered call system on potential customers compared to human agent calls. Leveraging a randomized field experiment conducted at a call center of a large securities company and a randomized online experiment, we investigated the mechanism resulting in the different impacts on customer behavior between humans and AI. The results show that voice-based AI calls trade off emotional and informational support: AI's informational advantages can raise intention, but empathy gaps can suppress it. These findings contribute to the literature on the application of technology in organizations and provide guidance to organizations on the effective implementation of AI systems, highlighting both the advantages and limitations of AI in customer-facing roles.

Deep learning (DL), as a vital technique, has sparked a notable revolution in AI, resulting in a great change in human lifestyles. As one of the most representative DL techniques, the Transformer architecture has empowered numerous advanced models, especially the large language models (LLMs) that comprise billions of parameters, becoming a cornerstone in deep learning. Despite the impressive achievements, Transformers still face inherent limitations, particularly the time-consuming inference resulting from the quadratic computation complexity of attention calculation. Recently, a novel architecture named Mamba , drawing inspiration from classical state space models (SSMs), has emerged as a promising alternative for building foundation models, delivering comparable modeling abilities to Transformers while preserving near-linear scalability concerning sequence length. This has sparked an increasing number of studies actively exploring Mamba’s potential to achieve impressive performance across diverse domains. Given such rapid evolution, there is a critical need for a systematic review that consolidates existing Mamba-empowered models, offering a comprehensive understanding of this emerging model architecture. In this survey, we therefore conduct an in-depth investigation of recent Mamba-associated studies, covering three main aspects: the advancements of Mamba-based models , the techniques of adapting Mamba to diverse data , and the applications where Mamba can excel . Specifically, we first review the foundational knowledge of various representative deep learning models and the details of Mamba-1&2 as preliminaries. Then, to showcase the significance of Mamba for AI, we comprehensively review the related studies focusing on Mamba models’ architecture design, data adaptability, and applications. Finally, we present a discussion of current limitations and explore various promising research directions to provide deeper insights for future investigations.

Large language models (LLMs) have recently been adopted for recommendations due to their ability to understand user intent and item semantics. However, LLM-based recommender systems often rely on parametric knowledge and suffer from outdated knowledge, motivating knowledge graph retrieval-augmented generation (KG-RAG) to ground recommendations on structured, up-to-date KGs. Despite this promise, effective KG-RAG in recommendations faces great challenges. First, users' queries vary in complexity and require KG knowledge at different granularities, whereas existing methods adopt a one-size-fits-all retrieval strategy, leading to over-retrieval for simple queries and under-retrieval for complex ones. In addition, augmenting LLMs with KG knowledge requires translating graph-structured data into linear text, which may introduce noise and cause structural information loss. Moreover, the selection of retrieval granularity lacks direct supervision and must be inferred from the final recommendation after alignment and downstream utilization, making query-aware retrieval hard to learn end-to-end. To address these issues, we propose MixRAGRec, a cooperative multi-agent framework for KG-RAG recommendations. MixRAGRec integrates a Mixture-of-Experts Retrieval Agent that routes each query to a KG retrieval expert with different granularities, a Knowledge Preference Alignment Agent that converts structured knowledge into LLM-friendly natural language, and a Contrastive Learning-reinforced Recommendation Agent trained with contrastive preference feedback. Notably, we introduce Mixture-of-Experts Multi-Agent Policy Optimization (MMAPO) to train three agents under a unified objective. Extensive experiments on real-world datasets demonstrate the effectiveness of our framework.

Purpose Visually impaired users face significant digital exclusion due to inaccessible interfaces, perpetuating barriers in education, employment and social participation. While assistive interfacing technology (AIT) is pivotal for accessibility, its design often misaligns with the capabilities of the visually impaired, culminating in high abandonment rates. This study reconceptualizes AIT design through affordance theory to bridge the gap between functional features and effective usage. Design/methodology/approach Based on a systematic literature review, we applied affordance theory to synthesize empirical insights regarding operational input and informational output of AIT. Design principles were derived from diverse prescriptions of AIT in past studies to reveal how operational input and informational output should be designed to maximize accessibility. Findings Focal principles in terms of operational input include context-aware modality selection, minimalist reusable actions and automated action sequencing to reduce cognitive load. Likewise, focal principles of informational output include multimodal redundancy, proactive granular disclosure and precision-calibrated feedback. Essentially, sensory compensation is essential in designing both operational input and informational output of AIT. We pioneer an affordance-centered framework that explicitly connects designer features of AIT with the capabilities of visually impaired users. Originality/value We refine theory by foregrounding nonvisual perception dynamics and delivering theoretically grounded insights to guide disability-centered human–computer interaction. Practitioners would benefit from our work by gleaning actionable principles that can be incorporated into the design of AIT to maximize accessibility to digital resources for the visually impaired.

PurposeThis study aims to reveal an inverted U-shaped relationship, showing that a moderate number of elite reviews stimulate the generation of subsequent peer reviews in both volume and novelty, while an excessive number of elite reviews inhibit it.Design/methodology/approachWe empirically support our framework using the Yelp Academic Dataset, which includes 500,013 reviews of 2,533 restaurants from 2004 to 2022. We use text mining to convert review texts into quantitative indexes and apply the Heckman two-stage model and propensity score matching to address potential endogeneity issues with robust results. Additionally, we employ GuidedLDA to classify restaurants into four types, highlighting the heterogeneity of our findings.FindingsWhile elite endorsements motivate peers to contribute, excessively increasing elites can be seen as manipulative, discouraging customer opinions. We also identify an inverted U-shaped pattern in elite impact on subsequent peer review novelty, where increasing elites limit topics, hindering novelty.Research limitations/implicationsFuture research should assess our findings' generalizability in other geographical locations and cultures.Practical implicationsOur findings help managers find the right balance in elite engagement, maximizing the benefits of elite recommendations while avoiding elite saturation and audience fatigue. Managers should carefully assess their elite recommendation strategies, emphasizing trust, authenticity and individuality within their audience. This ensures companies do not over-rely on elite recommendations, leading to a more effective and sustainable strategy.Originality/valueThis study addresses an identified need; that is, how elites' past opinions influence peers' future opinions, uncovering the potential inhibition effects of elites on the quantity and novelty of peer-generated content.

Nowadays, smart delivery tracking systems that combine GPS tracking and real-time logistics locations have enabled both sellers and consumers to see the exact delivery lead time. However, in a cross-border e-commerce system where delivery time seems more important for overseas consumers, we observe that many e-tailers have only quoted the delivery time based on expectation, resulting in a huge promised delivery time (PDT) cost, especially when the realized delivery time (RDT) is longer than PDT so a penalty cost is incurred. In this paper, we investigate an e-tailer's strategic decision to adopt the smart delivery tracking system where one multinational firm (MNF) sells products through both its own retail subsidiary and the third-party e-tailer. Clearly, the smart delivery tracking system allows the stakeholders to know the real-time RDT so the e-tailer saves the PDT cost. We reveal that there exists an overall-cost-mitigation effect in which the MNF is incentivized to lower the wholesale price to counteract the e-tailer's PDT cost. This drives the e-tailer's adoption of the smart delivery tracking system to switch twice, depending on the difference between the impacts of RDT and PDT on the online market potential. We further identify conditions under which the MNF also benefits from the e-tailer's adoption of the smart delivery tracking system, finding that the MNF may be beneficial, but consumer surplus may be hurt.

As one of the most representative DL techniques, Transformer architecture has empowered numerous advanced models, especially the large language models (LLMs) that comprise billions of parameters, becoming a cornerstone in deep learning. Despite the impressive achievements, Transformers still face inherent limitations, particularly the time-consuming inference resulting from the quadratic computation complexity of attention calculation. Recently, a novel architecture named Mamba, drawing inspiration from classical state space models (SSMs), has emerged as a promising alternative for building foundation models, delivering comparable modeling abilities to Transformers while preserving near-linear scalability concerning sequence length. This has sparked an increasing number of studies actively exploring Mamba's potential to achieve impressive performance across diverse domains. Given such rapid evolution, there is a critical need for a systematic review that consolidates existing Mamba-empowered models, offering a comprehensive understanding of this emerging model architecture. In this survey, we therefore conduct an in-depth investigation of recent Mamba-associated studies, covering three main aspects: the advancements of Mamba-based models, the techniques of adapting Mamba to diverse data, and the applications where Mamba can excel. Specifically, we first review the foundational knowledge of various representative deep learning models and the details of Mamba-1&2 as preliminaries. Then, to showcase the significance of Mamba for AI, we comprehensively review the related studies focusing on Mamba models' architecture design, data adaptability, and applications. Finally, we present a discussion of current limitations and explore various promising research directions to provide deeper insights for future investigations.

PurposeWe investigate the joint impacts of three trust cues – content, sentiment and helpfulness votes – of online product reviews on the trust of reviews and attitude toward the product/service reviewed.Design/methodology/approachWe performed three studies to test our research model, presenting participants with scenarios involving product reviews and prior users' helpful and unhelpful votes across experimental settings.FindingsA high helpfulness ratio boosts users’ trust and influences behaviors in both positive and negative reviews. This effect is more pronounced in attribute-based reviews than emotion-based ones. Unlike the ratio effect, helpfulness magnitude significantly impacts only negative attribute-based reviews.Research limitations/implicationsFuture research should investigate voting systems in various online contexts, such as Facebook post likes, Twitter microblog thumb-ups and up-votes for article comments on platforms like The New York Times.Practical implicationsOur findings have significant implications for voting system-providers implementing information techniques on third-party review platforms, participatory sites emphasizing user-generated content and online retailers prioritizing product awareness and reputation.Originality/valueThis study addresses an identified need; that is, the helpfulness votes as an additional trust cue and the joint effects of three trust cues – content, sentiment and helpfulness votes – of online product reviews on the trust of customers in reviews and their consequential attitude toward the product/service reviewed.

Recently, Large Language Model (LLM)-empowered recommender systems (RecSys) have brought significant advances in personalized user experience and have attracted considerable attention. Despite the impressive progress, the research question regarding the safety vulnerability of LLM-empowered RecSys still remains largely under-investigated. Given the security and privacy concerns, it is more practical to focus on attacking the black-box RecSys, where attackers can only observe the system's inputs and outputs. However, traditional attack approaches employing reinforcement learning (RL) agents are not effective for attacking LLM-empowered RecSys due to the limited capabilities in processing complex textual inputs, planning, and reasoning. On the other hand, LLMs provide unprecedented opportunities to serve as attack agents to attack RecSys because of their impressive capability in simulating human-like decision-making processes. Therefore, in this paper, we propose a novel attack framework called CheatAgent by harnessing the human-like capabilities of LLMs, where an LLM-based agent is developed to attack LLM-Empowered RecSys. Specifically, our method first identifies the insertion position for maximum impact with minimal input modification. After that, the LLM agent is designed to generate adversarial perturbations to insert at target positions. To further improve the quality of generated perturbations, we utilize the prompt tuning technique to improve attacking strategies via feedback from the victim RecSys iteratively. Extensive experiments across three real-world datasets demonstrate the effectiveness of our proposed attacking method.

In an era of information explosion, recommender systems are vital tools to deliver personalized recommendations for users. The key of recommender systems is to forecast users' future behaviors based on previous user-item interactions. Due to their strong expressive power of capturing high-order connectivities in user-item interaction data, recent years have witnessed a rising interest in leveraging Graph Neural Networks (GNNs) to boost the prediction performance of recommender systems. Nonetheless, classic Matrix Factorization (MF) and Deep Neural Network (DNN) approaches still play an important role in real-world large-scale recommender systems due to their scalability advantages. Despite the existence of GNN-acceleration solutions, it remains an open question whether GNN-based recommender systems can scale as efficiently as classic MF and DNN methods. In this paper, we propose a Linear-Time Graph Neural Network (LTGNN) to scale up GNN-based recommender systems to achieve comparable scalability as classic MF approaches while maintaining GNNs' powerful expressiveness for superior prediction accuracy. Extensive experiments and ablation studies are presented to validate the effectiveness and scalability of the proposed algorithm. Our implementation based on PyTorch is available.

Manipulated reviews can mislead consumers to make inappropriate purchase decisions and reduce consumers’ dependency on online reviews, jeopardizing the platforms’ reputation. Existing studies mainly focus on the detection and the impact of manipulated reviews but are limited in examining the determinants of review legitimacy from the perspective of consumer perception. The current research introduces perceived review manipulation, defined as the extent to which an individual perceives a review as non-authentic with the goal of misleading others and influencing product sales. Using psychological reactance theory as a general framework, we investigate the impact of reviews with deviation from the average ratings on perceived review manipulation and review adoption. We future examine two boundary conditions of the above relationship—the moderating effects of review content concreteness and reviewer rating distribution. We adopt a multi-method approach to the empirical test of the research model. First, three online randomized experiments reveal that: (1) reviews with deviant ratings are more likely to be perceived as manipulated; (2) the relationship in (1) is enhanced when review content is abstract rather than concrete and when a reviewer is usually negative/positive (i.e., his/her rating distribution has positive/negative skewness) based on the deviation direction; and (3) perceived review manipulation negatively influences review adoption. Second, a field study was conducted to support the external validity of our research model. Our findings have academic and practical implications.

Shipment consolidation is an effective way to reduce carbon emissions, because it can better utilise the ship-borne space and lower the shipment frequency. However, there also exists the dark side of shipment consolidation, i.e. the increasing pricing power of the common shipping company, and the ‘green dilemma’ because of green demand resluted by the shipping company’s emission reduction efforts. This paper studies the shipment consolidation decision of a retailer giant who resells two substitutable products, where either a common shipment company or two exclusive shipping companies can be contracted. The main findings include: (1) shipment consolidation induces the common shipping company to invest more in emission reduction, but strengthens the green dilemma, worsening environmental performance. (2) Pareto improvement of economic and environmental sustainability can be achieved with shipment consolidation. (3) Shipment consolidation can be even beneficial for upstream suppliers, so an ‘all-win’ situation is possibly observed.

Purpose This study clarifies the integration-related effects of photos and text on consumer information processing and decision-making outcomes. Design/methodology/approach The authors conducted an experiment by recruiting 162 workers from Amazon Mechanical Turk. These participants were randomly assigned based on a full factorial, between-subject design with four possible conditions (2 [separate vs alternate layout] × 2 [photo-first vs text-first sequence]). The authors conducted a two-way analysis of variance to test the main effects and the interaction effects of layout and sequence on perceived diagnosticity, pleasantness feelings and attitudes toward products or services reviewed through electronic word-of-mouth (e-WOM); the authors also applied Process Models 4 and 8 to explore the mechanism of these effects. Findings The experimental results reveal that text-first sequence is generally more effective than photo-first sequence in enhancing perceived diagnosticity and attitudes toward products or services. However, when a photo is displayed first, a separate layout is more effective than an alternate layout in enhancing perceived diagnosticity and attitudes. By contrast, regardless of the sequence, an alternate layout is more effective than a separate layout in inducing pleasantness feeling. Research limitations/implications Future studies should further explore photo-based e-WOM, including other photo characteristics (e.g. visual quality, quantity and content). Practical implications This study provides guidelines for businesses to use photos on social media to achieve strategic goals. Originality/value This study addresses an identified need; that is, how the presentation of photo cues (e.g. layout and sequence) influences consumer decisions.

The COVID-19 pandemic has underscored the urgent need for healthcare entities to develop resilient strategies to cope with disruptions caused by the pandemic. This study focuses on the digital resilience of certified physicians who adopted an online healthcare community (OHC) to acquire patients and conduct telemedicine services during the pandemic. We synthesize the resilience literature and identify two effects of digital resilience-the resistance effect and the recovery effect. We use a proprietary dataset that matches online and offline data sources to study the digital resilience of physicians. A difference-in-differences (DID) analysis shows that physicians who adopted an OHC had strong resistance and recovery effects during the pandemic. Remarkably, after the COVID-19 outbreak, these physicians had 35.0% less reduction in medical consultations in the immediate period and 31.0% more bounce-back in the subsequent period as compared to physicians who did not adopt the OHC. We further analyze the sources of physicians' digital resilience by distinguishing between new and existing patients from both online and offline channels. Our subgroup analysis shows that, in general, digital resilience is more pronounced when physicians have a higher online reputation rating or have more positive interactions with patients on the OHC platform, providing further support for the mechanisms underlying digital resilience. Our research has significant theoretical and managerial implications beyond the context of the pandemic.

Discovering the most adaptive learning path and content is an urgent issue for nowadays e-learning environment, for achieving learning goals efficiently and effectively. The main challenge of building this system is to provide appropriate educational guide and resource for different learners with respective interests and knowledge base. In order to reduce people's cognitive overload and fulfill their self-learning requirements, this article proposes a framework for a self-learning system. The system is design to be closed and updated automatically, in which learning path is discovered based on differential evolution (DE) algorithm and knowledge graph. The output of the system includes: (1) the personalized learning path adapted to learner's specific needs; (2) learning resource recommendation matching the learning path; (3) test results of learners' learning effect after following the learning path and resources recommendation; (4) revised learning path and resources recommendation according to learner's evaluation. Experimental results show that the system based on DE algorithm and disciplinary knowledge graph is feasible in optimal learning path discovery and further learning resources recommendation.

Real-time payment (RTP) enabled by FinTech is changing the cross-border B2B transactions that have frequently suffered from delay payment issues. In this paper, we study a supply chain where a multinational firm (MNF) produces products in its domestic manufacturing division and then sells them to overseas customers through both its retailing division and a local retailer. We investigate the supply chain parties' preferences for RTP by considering the trade-offs among the MNF's tax-planning, the cash opportunity cost, and the downstream competition between the MNF's retailing division and the local retailer. We find that the retailing division and the local retailer always have conflicts of interest facing RTP so their preferences cannot be aligned. For the MNF, interestingly, we find that RTP can be preferred even if the MNF's relative cash opportunity cost is large, depending on the tax disparity between its manufacturing and retailing divisions. We also find that tense local competition will motivate the MNF to prefer RTP, especially when the tax disparity is sufficiently small.

Purpose A smart city is a potential solution to the problems caused by the unprecedented speed of urbanization. However, the increasing availability of big data is a challenge for transforming a city into a smart one. Conventional statistics and econometric methods may not work well with big data. One promising direction is to leverage advanced machine learning tools in analyzing big data about cities. In this paper, the authors propose a model to learn region embedding. The learned embedding can be used for more accurate prediction by representing discrete variables as continuous vectors that encode the meaning of a region. Design/methodology/approach The authors use the random walk and skip-gram methods to learn embedding and update the preliminary embedding generated by graph convolutional network (GCN). The authors apply this model to a real-world dataset from Manhattan, New York, and use the learned embedding for crime event prediction. Findings This study's results show that the proposed model can learn multi-dimensional city data more accurately. Thus, it facilitates cities to transform themselves into smarter ones that are more sustainable and efficient. Originality/value The authors propose an embedding model that can learn multi-dimensional city data for improving predictive analytics and urban operations. This model can learn more dimensions of city data, reduce the amount of computation and leverage distributed computing for smart city development and transformation.

B2B marketers are encouraged to leverage UGC in accomplishing marketing goals, yet have serious concerns in encouraging their clients to post UGC and in leveraging it. In drawing from in-depth interviews with B2B practitioners and building on previous research on information technology and B2B marketing, our findings and analysis address these concerns with confidentiality and conflicts of interest by relating them to the B2B characteristics of multiple stakeholders, high value exchanges, complex and elaborate specifications, and long-term relationships. Extrapolating from these findings and analysis, discussion presents a framework elaborating the multi-organizational information system of B2B UGC that aids B2B marketers and researchers in deciphering the impacts of B2B characteristics on marketers' efforts to encourage clients to generate UGC, and to leverage such UGC. Contributions specify the inter-organizational information system of B2B UGC, add precision in defining B2B UGC marketing, and add to knowledge of the impacts of B2B characteristics on B2B UGC generation and leverage, while recommendations guide B2B marketers in addressing their concerns in ways that generate and leverage UGC towards the goal of market development for the benefit of all collaborators.