Large language models are increasingly deployed in safety critical domains, yet they remain vulnerable to jailbreak attacks that bypass alignment safeguards through carefully constructed prompts. We propose JAIL, a multi-turn jailbreak framework that decomposes a malicious objective into a sequence of conversational queries and gradually steers a target model toward policy violating outputs. JAIL formulates jailbreak generation as an interactive optimization problem. At each turn, an attacker model generates candidate queries conditioned on the dialogue history and the ultimate objective. A loss guided beam search procedure selects queries that maximize estimated attack effectiveness while maintaining linguistic fluency, enabling systematic escalation across turns. To improve stability and sample efficiency, we incorporate Direct Preference Optimization, which allows the attacker to learn from both successful and unsuccessful trajectories through query level preference signals. We evaluate JAIL on a range of open source and API based large language models. Across models and evaluation metrics, JAIL consistently outperforms strong single turn and multi-turn baselines, achieving substantially higher attack success rates under limited interaction budgets. These results expose structural limitations in existing alignment mechanisms and suggest the need for defense strategies that account for adversarial dialogue dynamics. Code is provided in the supplementary material.
As a workplace stressor, illegitimate tasks exert adverse impacts on employees in various ways. While recent studies have begun to explore the dual nature of illegitimate task and reveal its opposing effects on the same outcome, it remains unclear under what conditions and through which mechanisms the task produce divergent behavioral responses among employees. This research examines the dual potential of illegitimate task through the lens of cognitive appraisal theory of stress. To test the suggested conceptual model, we gathered and analyzed the data of 275 employees across six enterprises in southwestern China. The findings indicate that employees high in need for achievement often utilize approach coping strategies when encountering illegitimate tasks, thereby facilitating proactive behavior. Contrarily, those low in need for achievement tend to employ avoidance coping strategies in the face of such tasks, which result in procrastination behavior. Theoretical and practical implications are provided.
As an effective reliability management tool, failure mode and effect analysis (FMEA) has been widely utilized in various fields. However, existing studies overlook the fact that words may mean discrepant things to different people, which entail that FMEA participants have personalized individual semantics (PISs) in terms of their linguistic expressions. Moreover, decision-makers (DMs) may have different requirements or restrictions on risk factors in FMEA. This study attempts to develop a novel multi-criteria group decision support framework for FMEA with linguistic distribution assessments considering PISs from the perspective of tolerance attitudes. First, the concept of discrimination measure is defined, facilitating the formulation of a discrimination-based optimization model for assigning personalized numerical scales to team members. Second, several consensus-based optimization models are designed to provide adaptive adjustment strategies to yield consensual solutions. Third, the local and global tolerance attitudes of DMs to risk factors are fully modeled into the multi-criteria evaluation process through 2-addtive Choquet integral and investigate the impact mechanism of the risk tolerance attitudes of DMs on the final results. Finally, a case study on electric vehicle business model innovation with sensitive and comparative analysis is presented to verify the application of the proposed framework.
The digital revolution has fundamentally transformed banking through financial technology (FinTech). We examine whether bank FinTech can enhance innovation capabilities among Chinese small- and micro-sized enterprises (SMEs). Using web scraping technology, we construct a novel city-level bank FinTech index. Our findings demonstrate that bank FinTech significantly boosts SME innovation compared to traditional branch expansion. Mechanism analysis reveals that FinTech’s innovation incentive effects primarily operate through two channels: alleviating financing constraints and reducing rent-seeking costs. However, we identify a “universal but not beneficial” phenomenon, as FinTech fails to significantly lower credit interest rates. Notably, the innovation incentive effects are particularly pronounced in regions with strong inclusive finance policy implementation. Heterogeneity analysis indicates that bank FinTech most effectively promotes innovation both among SMEs lacking physical capital and among those located in eastern China, while showing limited impact on those with low human capital. These findings provide valuable policy implications for stimulating SME innovation and development in the digital era. Bank FinTech boosts SME innovation more effectively than traditional bank branches! We show that bank FinTech helps small businesses (i.e., SMEs) innovate by easing credit access and cutting rent-seeking costs—though it doesn’t lower interest rates. These effects are strongest in regions with strong inclusive finance policies, in eastern China, and for SMEs lacking physical capital but with high human capital. Additionally, bank FinTech helps SMEs pursue higher-level innovation and upgrade their human capital. For policymakers, this means FinTech adoption, paired with inclusive finance reforms, can be a powerful tool to spur innovation, especially for underserved firms. Small businesses should fully utilize bank FinTech to overcome funding barriers, while researchers should explore how to make FinTech benefits more widespread.
This study examines the impact of China’s 2019–2023 climate policies on the stock returns of sectors within the carbon-neutral industry chain. By employing contagion tests and network analysis, we explore how policy shocks propagate from the carbon-neutral sector to its upstream, midstream and downstream sectors. Our analysis also quantifies the importance of climate policy in driving policy shocks and identifies the main driver within the carbon-neutral industry chain. The empirical findings indicate that the 2019 persuasive strategy policy demonstrates the most robust evidence of policy contagion, followed by the 2020 cap-and-trade scheme and the 2022 and 2023 carbon offset mechanisms. In contrast, the regulatory measures exhibit the weakest evidence of policy contagion. Notably, climate policies have a greater impact on upstream sectors, which consist of carbon-intensive industries, compared with the midstream and downstream sectors, which comprise renewable energy sectors.
