Large language models (LLMs) are increasingly used for decision support and behavioral simulation, yet it remains unclear whether they reproduce classic mental accounting biases and how these patterns depend on model configuration. This study examines whether LLMs exhibit hedonic framing, account non-fungibility, and sunk cost sensitivity, and how these tendencies respond to demographic personas and sampling temperature. We evaluate Gemini 2.0 Flash, Claude 3.5 Sonnet, GPT 4o, and DeepSeek V3 on a curated set of canonical decision scenarios that are also administered to 306 human participants in text-only tasks. Each item is replicated across multiple independent runs with majority aggregation, and we estimate prospect theory parameters and fit regression models to compare decision patterns across humans and models. The results show that LLMs often mirror human-like mental accounting, including segregating gains and losses, treating nominally equivalent resources as non-fungible, and honoring sunk costs, although the strength and direction of these effects differ across models and conditions. Demographic personas and temperature systematically modulate these biases, shifting model behavior between more rational and more human-biased regimes without eliminating mental accounting effects. These findings clarify when LLMs can approximate human decision behavior in economic tasks and provide practical guidance for configuring LLM-based agents in applications such as project continuation, pricing, and portfolio choices.
In large-scale group decision-making (LGDM), experts often exhibit behavioral heterogeneity arising from diverse cognitive patterns, risk attitudes, and willingness to cooperate, which poses significant challenges to achieving group consensus. To address this issue, this study develops a behavioral heterogeneity-oriented (BHO) consensus model in a non-cooperative environment, aiming to balance expert diversity with the overall level of collective consensus. First, personalized adjustment costs are constructed by measuring experts’ degrees of hesitation and trust-related risk attitudes, and a minimum cost consensus (MCC) model is employed to generate individualized opinion recommendations, thereby facilitating consensus while respecting experts’ behavioral heterogeneity. Second, non-cooperative behavior (NCB) is identified based on the deviation between experts’ actual adjusted opinions and the recommended adjustments. In this process, considering that experts exhibit specific behavioral patterns during preference adjustment, an adjustment deviation measure based on the technique for order preference by similarity to an ideal solution using a modified Mahalanobis distance (M-TOPSIS) is proposed to capture the correlations and distributional characteristics of adjustment behaviors. Meanwhile, a cumulative effect mechanism is introduced to further amplify repeated excessive deviations and adaptively update expert weights, thereby mitigating persistent non-cooperative adjustments. Finally, an application example and simulation analyses demonstrate that incorporating behavioral heterogeneity into the consensus-reaching process (CRP) can effectively reduce consensus costs and enhance consensus levels, confirming the effectiveness of the proposed approach.
As Artificial Intelligence (AI) assumes a growing presence in executive decision spaces, an empirical and theoretical gap persists concerning the boundary conditions governing how top management teams (TMTs) integrate AI into strategic decision-making (SDM). Existing research addresses AI’s impact on operational, functional, and middle-management decisions, yet leaves the unstructured terrain of Top management teams’ strategic decision-making largely unexplored. We argue that SDM at the apex of organisations is distinct not because any single attribute is unique to it, but because four attributes such as the irreversibility of strategic decisions, the socio-political collectivity of TMT processes, the proprietary character of executive AI tools, and the identity stakes of strategic leadership may co-occur and reciprocally amplify each other, producing integration dynamics qualitatively different from those theorized in the broader human-AI collaboration literature. We address two research questions: (1) Under what contextual conditions do TMT members delegate strategic decisions to AI rather than treat AI as an augmentation resource? (2) How does the structure of strategic decisions and the predictability of their situational context shape the mode and depth of human-AI collaboration in TMT practice? Guided by the Gioia methodology and drawing on 33 semi-structured interviews with C-suite executives across three multinational organisations, supplemented by structured demonstrations of in-house AI tools at each firm, we develop a grounded framework distinguishing high-confidence AI delegation of analytical work from iterative human-AI co-deliberation along two axes: problem structuredness and situational predictability. Eighteen first-order codes, six second-order themes, and two aggregate dimensions as (a) Contextualized AI Integration and Delegation Framework, and (b) the Transformative Human-AI Strategic Empowerment dimension, emerge from systematic inductive analysis. Three theoretically grounded propositions offer implications for Upper Echelons Theory, Strategic Decision-Making Theory, and Contingency Theory, providing a TMT-specific account of human-AI integration that goes beyond the generic automation–augmentation divide.
In transboundary river basins, achieving spontaneous cooperation over water between upstream and downstream nations is often challenging due to conflicting interests under water scarcity, especially in the absence of intergovernmental water cooperation organizations (IWCOs). Existing studies frequently assume IWCOs act as strong reciprocators, yet they often overlook the associated regulatory costs, as well as the coexistence of conflict and cooperation among riparian countries. This study develops a tripartite game model that incorporates upstream nations as water suppliers, downstream nations as water users, and an IWCO as the regulator. By integrating factors such as water supply benefits, water utilization benefits, speculative gains, and regulatory measures, the model is applied to the Lancang–Mekong River Basin. The results indicate that: First, the IWCO acts as a cost-sensitive “catalyst”; high supervision costs may lead it to adopt “weak regulation,” yet cooperation can still spontaneously emerge when driven by sufficient shared interests. Second, conflict and cooperation are not mutually exclusive but co-evolve across distinct phases; early-stage conflict acts as a functional signal triggering necessary intervention, which eventually transitions into stable cooperation sustained by later-stage benefit-sharing. Third, stability is threatened by speculative micro-conflicts, which can be suppressed through a precise mechanism combining targeted penalties and incentives. This study provides a structured framework for achieving sustainable transboundary water management.
Dempster–Shafer evidence theory (D–S theory) is a well-known framework for reasoning under uncertainty. It can effectively combine information from multiple sources using Dempster’s rule of combination. However, when the information in the bodies of evidence (BOEs) is highly conflicting, applying Dempster’s rule directly can produce counterintuitive or unreliable results. To address this issue, we proposed a novel approach that models BOEs as nodes within a complex network. In this network, nodes (BOEs) were correlated with each other to varying degrees, quantified by classical evidence distance. The importance of each node was then determined by calculating its direct and indirect strength, both derived from these correlation degrees. Based on the strength measures, we assigned weight factors to adjust the belief degrees in the original evidence bodies. Finally, the weighted evidence was fused using Dempster’s combination rule. Experimental results demonstrated that our method effectively mitigates the limitations of the traditional rule, showing superior convergence speed and enhanced reliability when fusing highly conflicting evidence.
Artificial intelligence has been rapidly increasing in sophistication and impact, reinforcing discussions of its extensive potential to disrupt individual practices and organizational designs. We consolidate data from the O*NET 2023 database on job roles to create a list of 85 corporate knowledge worker roles and associated skill levels. Next, together with five different genAI applications, we assess the 15 low skill level roles to identify which roles genAI will automate or augment (human with genAI) in the near future, comparing the genAI analysis with the analysis of a human team. Hence, we do not just study the impact of genAI, we also team up with it. Our analysis shows that the increasing use of genAI disrupts how organizations function, with changes occurring at the individual level (which we theorize via job crafting theory), and over time impacting the way organizations function.
Large language models (LLMs) are increasingly used as decision-support tools in domains characterized by risk and uncertainty, raising fundamental questions about how their behavior aligns with normative standards of rational choice. In this study, we develop and apply a structured battery of Expected Utility Theory (EUT)-based decision tasks to evaluate three widely used frontier models, ChatGPT, Claude, and DeepSeek, under both deterministic and non-deterministic settings. The battery spans five canonical domains of EUT violation, covering certainty effects, reflection effects in losses, mixed gambles, rare-event valuation, and insurance parity. Across models and conditions, we find that departures from expected utility are systematic, repeatable, and strongly model-specific, producing distinct and stable bias profiles that persist even when stochastic variation is removed. These findings extend the concept of bounded rationality beyond human cognition to collaborative intelligence systems in which decision outcomes emerge from the interaction of human users and AI models. Rather than acting as neutral correctives, LLMs instantiate distinct and persistent forms of bounded rationality and may, in some contexts, amplify existing cognitive distortions. We argue that these results have direct implications for bias management in human–AI decision-making systems. Structured EUT violation batteries provide a practical tool for diagnosing model-specific biases and should become a standard component of model evaluation and deployment in high-stakes settings. More broadly, our findings support a shift from attempts at bias elimination toward systematic bias management, combining normative diagnostics, model-specific testing, and sustained human oversight.
Large-scale group decision-making problems based on social network analysis have garnered significant attention in recent years. Yet, few studies have addressed the challenge of expert heterogeneity in such decision-making. So, this paper constructs a large-scale group decision-making model based on a local search algorithm and heterogeneous feedback adjustment. Specifically, we employ a novel community detection method, the local search algorithm, to mine experts’ local information and identify community structures, constructing a multi-layer network model to clearly characterize the dynamic interaction relationships among experts. Building upon this foundation, we introduce a multi-attribute bounded confidence model. By integrating social network analysis metrics such as degree centrality and proximity centrality, we define both the global and local weights of experts. Additionally, to address the role differences between leaders and followers within the community, a consensus-reaching process with heterogeneous feedback mechanisms was designed; leaders focus on opinion steering, while followers concentrate on consensus building, thereby enhancing decision-making efficiency. In simulation analysis, we demonstrated that incorporating leader guidance, specifically, designing a heterogeneous feedback adjustment mechanism, can reduce the number of iterations required to reach consensus. Furthermore, as the leader’s weight increases, the number of iterations decreases. Finally, case studies and comparative experiments validated the effectiveness and practicality of the proposed method.
Scientific and democratic decision-making mechanisms require integrating the collective wisdom of multidisciplinary decision-makers (DMs), while increasingly complex real-world decision environments pose additional challenges to the consensus reaching process (CRP). This paper focuses on studying multi-attribute group decision-making methods that incorporate DMs’ bounded confidence and non-cooperative behaviors under incomplete information. First, the paper utilizes reference risk to help DMs with missing evaluation information complete the missing values. In the feedback mechanism during CRP, the study measures DMs’ adjustment willingness based on bounded confidence and their social trust relationships, proposing an objective method for calculating feedback parameters. Subsequently, the paper identifies and manages non-cooperative behaviors from both evaluation information and trust relationship perspectives. Finally, the effectiveness of the proposed method is validated through a case study of hot dry rock development site selection. The results of the comparative analysis illustrate the feasibility and innovation of the consensus decision-making method proposed in this paper.
Large Language Models (LLMs) are transforming the innovation process by scaling ideation and creativity. This yields the challenge of evaluating the growing flood of ideas. In simpler contexts, tasks with clear criteria and well-defined outcome possibilities, where human evaluators are also a bottleneck, LLM systems are already being used as a scalable and cheap alternative. In this study, we assess whether LLMs and LLM agents can be used as evaluators in innovation processes, making complex evaluations that require probabilistic judgments about future outcomes. Thereby, we examine the effect of different system, group, and collaboration architectures on the performance. Results indicate that implementing more LLMs instead of a single LLM is beneficial, traceable to wisdom of the crowd effects. Moreover, our results show that agent heterogeneity and structured collaboration provide added value, leveraging collective intelligence effects that further enhance evaluation performance beyond what stronger models achieve individually. Findings contribute to the theoretical understanding of LLM system architectures and discuss the practical relevance of LLM evaluators.
Multiple attribute group decision-making is widely adopted across numerous domains, with consensus-reaching processes commonly employed as the standard solution approach. Despite prior efforts to minimize the associated costs, the expense incurred from multiple rounds of discussions remains significant and escalates rapidly with increasing stakeholder group size. Stakeholder consultations and engagements have emerged as efficient surrogates during early-stage planning, where networked stakeholders influence each other’s voting preferences through offline communication. However, identifying the most influential stakeholders amid the growing complexity of larger projects remains a challenging task. This paper proposes a novel mathematical model inspired by the Ising Model from physics to represent the stochastic influence propagation within stakeholder networks. Subsequently, we present two algorithms designed to calculate the expected agreement rate and identify the optimal set of stakeholders, applicable to various stakeholder networks and voting criteria. Experimental results further demonstrate the effectiveness of our proposed algorithms in facilitating efficient stakeholder engagement.
Failure Modes and Effects Analysis (FMEA) is a proactive risk management tool used to identify key factors in system failures and mitigate associated risks. However, during FMEA implementation, experts from diverse backgrounds often express divergent opinions, and their interactions may not only influence the perspectives of others but also trigger non-cooperative behaviors. To address these challenges, this study proposes an FMEA classification consensus framework aimed at examining experts’ interactive and non-cooperative behaviors during the consensus reaching process (CRP), and categorizing failure modes (FMs) into several risk-ordered classes to support risk management. First, a minimum-adjustment best-worst method (BWM) is introduced to derive initial risk scores of FMs from individual experts. Subsequently, a maximum consensus model incorporating expert opinion interactions is established to determine expert weights and interaction indices. The degree of non-cooperation is then quantified based on these interaction indices, and experts are classified using a predefined consensus–noncooperation threshold. Furthermore, tailored feedback mechanisms are designed to guide experts in revising their opinions, thereby managing non-cooperative behaviors and promoting consensus achievement. Finally, the proposed method is applied to assess FM risks in the radiation therapy process, validating its feasibility and effectiveness.
Group decision-making and negotiation increasingly take place in settings where stakeholders hold divergent objectives, values, and interpretations of evidence. However, Large Language Models (LLMs) integration in collective decision processes remains constrained by limited traceability, weak procedural control, and ambiguity regarding the role of human judgment. This conceptual paper proposes a reference architecture for agentic Artificial Intelligence (AI) in Group Decision and Negotiation (GDN) that integrates language-based reasoning with formal Multi-Criteria Decision-Making (MCDM) procedures. The architecture assigns two complementary classes of specialized agents to discrete stages of the process: generative agents, responsible for interpretative tasks such as problem structuring, criteria definition, and preference elicitation, and logical agents, responsible for deterministic operations including weighting, aggregation, and ranking. A human-in-the-loop (HITL) governance layer supervises tasks requiring subjective judgment or domain expertise, ensuring consistency, transparency, and auditability throughout the decision workflow. The primary contribution is a modular reference architecture, grounded in design science principles, that decouples generative interpretation from formal evaluation within a unified and auditable decision pipeline. The framework is illustrated through a representative multi-stakeholder scenario demonstrating the coordination of agents and human oversight across all stages of the MCDM process.
This paper presents a novel framework for negotiation support using linguistic information. It employs the Linguistic Ideal-Based Method (LIBM) to evaluate and rank negotiation offers, linking linguistic terms to a numerical scale and using the Generalized Distance Measure (GDM2) to process ordinal data. By incorporating an ideal linguistic value, LIBM prevents rank reversal when offers are added or removed. Unlike conventional numerical approaches, which may distort subjective preferences, or fuzzy-based methods, which are cognitively demanding, LIBM reduces cognitive burden. It also aligns with users’ linguistic reasoning, making it accessible to negotiators with limited numerical skills. LIBM was compared with the Direct Rating (DR) method, highlighting its strengths and limitations. It was also analyzed alongside alternative linguistic approaches, including Measuring Attractiveness near Reference Situations (MARS) and fuzzy methods, demonstrating advantages in cognitive ease and practical usefulness. LIBM underpins the Linguistic Multi-Criteria Negotiation Support (LMNS) model, using a structured, multi-phase approach to enable transparent evaluation and decision guidance. An illustrative example shows LMNS’s effectiveness in simplifying evaluation, reducing cognitive effort, and ensuring practical relevance across negotiation contexts. These features position LMNS as a cognitively grounded, user-friendly tool that supports both symmetric and asymmetric scenarios, providing a robust alternative to numerical and fuzzy approaches.
Wisdom of the crowds (WoC) has been extensively studied as an approach to generate better solutions to solve a wide range of problems, such as financial predictions, marketing, and management decision-making. Nonetheless, there seems to be a lack of clarity about how it works. To better understand how WoC is conceptualized and can be applied, and to create a solid foundation for its development, we have reviewed the WoC and crowd performance literature. Based on our analysis, we: (1) integrate dispersed findings into an Input–Process–Output(IPO)-based multi-stage process model that links preparation inputs to judgment generation and aggregation processes and, ultimately, to the outputs—the manifestation of crowd performance; (2) explain why commonly presumed WoC “enabling” conditions yield mixed effects and translate this logic into three testable propositions: a Goldilocks (non-linear) crowd-size effect, shared biases or common cues as boundary conditions that weaken the benefits of diversity, independence, and decentralization, and a cost-effectiveness trade-off between accuracy and overall value; and (3) conceptualize and define crowd performance and distinguish it from related terms to reduce conceptual ambiguity. The practical implications and directions for future research are also provided.
The use of online reviews for ranking products has gained increasing interest in recent years. Yet, previous research has mainly single data modal in online reviews and do not consider consumer heterogeneity. Furthermore, the heterogeneity of consumers requires product ranking to be more targeted. To address these problems, this study proposes a multimodal data-driven personalized decision-making model for supporting purchase decisions of heterogeneous consumers. First, multimodal data (i.e., unstructured textual review data and structured ratings data) is crawled and preprocessed, then a new sentiment analysis process based on prospect theory is proposed. Second, a multimodal data fusion process is developed, which consists of ratings compression and information fusion. Third, this study presents a product ranking model for heterogeneous consumers. In this ranking model, heterogeneous consumers are divided into three categories according to consumers’ familiarity with the product, three methods related to criteria weight determination are developed to depict the personalization of heterogeneous consumers, and regret theory is used to characterize consumers’ decision-making psychology. Finally, a case study on helping consumers purchase new energy vehicles (NEVs) is provided to verify the feasibility and effectiveness of the product ranking model. In addition, the results of the benchmark analysis, the sensitivity analysis and the comparative analysis indicate that our personalized decision-making model is robust and effective for supporting purchase decisions of heterogeneous consumers.
With the swift advancement of social technology, the problem of Large-Scale Group Decision Making (GDM) has emerged as a new research focus. Nevertheless, upon conducting an in-depth exploration, it is discerned that this field is beset with many challenges. For instance, conventional clustering techniques struggle to cope with the complex structures of large-scale data, and group manipulation behavior in the feedback adjustment stage can seriously disrupt the fairness of the decision-making process. Therefore, this paper advances a large-scale multi-attribute GDM method that integrates the Gaussian mixture model (GMM) and is efficacious in preventing group manipulation. Firstly, the GMM is incorporated into the clustering of probabilistic linguistic term sets. Subsequently, the probability information gleaned from the clustering outcomes is utilized to generate the weights of decision-makers, which are then aggregated to derive the comprehensive perspectives of the distinct subgroups. Thereafter, the discrepant subgroup opinions that surface during the consensus measurement are subjected to adjustment. Considering the group manipulation issue that may arise during this adjustment, we establish an optimal feedback model with the minimization of adjustment cost and the maximization of fairness level to prevent group manipulation. Ultimately, the effectiveness and superiority of the proposed approach are affirmed by a low-altitude economy numerical analysis.
This paper first proposes the combination of preference-approval structures with ranking opportunity sets (i.e., collections of options), a relevant topic in welfare analysis. We suggest that various possibilities emerge, and we illustrate them with specific examples in finite contexts. Characterizations of three rankings of opportunity sets defined from preference-approval structures on the individual options are proven. Their representability by utilities is considered too. Various preference-approval structures on opportunity sets are defined from the information contained in a preference-approval structure on the options, that supplement these rankings of opportunity sets with appropriate collections of “approved” opportunity sets. Finally, preference-approval structures on opportunity sets are defined using the information derived from a ranking of the options. As a secondary contribution, we compute the number of preference-approval structures on a set with finite cardinality, thereby enriching the fundamental theory of this enhanced type of preferences.
This study examines the relationship between task conflict and relationship conflict in organizational settings. Drawing on attribution theory, it considers how the same disagreement may be construed in task-related or relational terms. The study aims to examine whether and how these forms of conflict are associated in ways consistent with a transformation pattern, and how these patterns relate to work outcomes. Data were collected from 268 employees across nine industries and analyzed using structural equation modeling. The findings indicate a positive association between task and relationship conflict in both directions, suggesting that these conflict types are interconnected rather than independent categories. Task type conditions this association, such that the link between relationship conflict and task conflict is stronger in non-routine contexts. The results suggest that the relationship between conflict and job satisfaction depends on the interconnections among conflict types rather than on the presence of a single conflict type in isolation. Task conflict is indirectly associated with job satisfaction through its association with relationship conflict, whereas the reverse pathway appears more conditional. Although attributional processes are not tested directly, the findings are consistent with an attribution-based interpretation of conflict dynamics. By adopting a process-oriented perspective, the study contributes to the conflict and negotiation literature by reframing task and relationship conflict as related and by clarifying how their interrelationship relates to job satisfaction.
This research explores how anger expression operates in structured group decision-making, a context that differs in important ways from the dyadic bargaining settings that dominate prior research. Using spatial voting theory and experimental designs informed by political science, we examined the effect of anger expression in decision-making groups with bases of power grounded in agenda control and voting position. Early experiments with these forms of power suggested gender differences in reactions that we also examine systematically. Through three experiments (two with programmed groups and one with in-person groups) conducted in the United States with English-speaking participants, we observed that anger expressions failed to consistently yield influence. They sometimes provoked retaliation, particularly among female participants in empowered roles. We proposed and tested the mechanisms behind the retaliation effect from the perspective of the EASI framework. Theoretical and practical implications are discussed.
