Large language models (LLMs) are increasingly deployed in highly sensitive domains such as education and career guidance, raising concerns about their potential to reproduce and amplify social biases. The present research examined whether LLMs exhibit gendered empathy stereotypes—specifically, the belief that “women are more empathetic than men”—and whether such stereotypes influence downstream recommendations. Three studies were conducted. Study 1 compared LLMs with human participants and found that across six leading LLMs, gendered empathy stereotypes were significantly stronger than those observed in humans across three facets of empathy: emotional empathy, attention to others’ feelings, and behavioral empathy. Study 2 manipulated input language (Chinese vs. English) and gender-identity priming (male vs. female), demonstrating that English prompts and female priming elicited stronger gendered empathy stereotypes. Study 3 focused on major and career recommendation tasks and revealed that LLMs systematically recommended high-empathy majors and professions to women, while directing men toward low-empathy fields. Together, these findings indicate that LLMs exhibit pronounced gendered empathy stereotypes, that these biases vary across input context, and that they can transfer into real-world recommendation scenarios. This research offers theoretical insights into bias formation in LLMs and provides practical implications for improving fairness in AI systems used in educational and career guidance.

Risky choice (RC) is a common and important form of decision making in daily life. Its theoretical development primarily follows two major theories: normative theory and descriptive theory. The paradigms of single- and repeated-play gambles can provide an effective framework for distinguishing between the theories. However, prior research lacks direct observations of the decision-making process, which can limit the deep understanding of individual behaviour and hinder the development of effective behavioural interventions. In recent years, large language models (LLMs) have demonstrated highly human-like characteristics by not only simulating human preferences in behavioural performance but also exhibiting similar reasoning pathways. This offers a promising solution to the aforementioned limitations. This study, which is grounded in the classic RC paradigms of single versus repeated gambles, investigates the capability of LLMs to simulate and understand risk preferences and decision-making processes. Specifically, this study explores the potential of LLMs’ understanding of decision strategies to generate intervention texts and evaluates their effectiveness in influencing human decisions.

This work comprises three studies. In Study 1, GPT-3.5 and GPT-4 were employed to simulate human responses to gambling decisions under nine probability conditions (with constant expected value), which generated a total of 3, 600 responses across single and repeated gamble scenarios. In Study 2, LLM-generated strategies were constructed through a three-stage process (decision rationale extraction, strategy generation and quality evaluation), then the human participants were required to complete decision-making tasks in two experiments: Experiment 1 replicated the medical/financial scenarios (N = 349, N _male = 174, M _age = 21.79) of Sun et al. (2014) in a 2 (context: medical vs. financial) × 2 (application frequency: single vs. repeated) within-subjects design, and Experiment 2 examined digital contexts with a 2 (context: content creation vs. e-commerce marketing) × 2 (frequency: single vs. repeated) mixed design (context as between subjects). Subsequently, DeepSeek-R1 was used to perform the same tasks and generate strategy texts through the three-stage process. Finally, the participants were instructed to evaluate their acceptance of the LLM-generated strategies. Study 3 extended the Study 2 methodology to determine whether the LLM-generated intervention texts could reverse the participants’ classic choice preference across the single versus repeated gamble scenarios. The Study 2 experimental contexts (Experiment 1: medical vs. financial, N = 460, N _male = 205, M _age = 21.80; Experiment 2: content creation vs. e-commerce marketing, N = 240, N _male = 106, M _age = 29.12) were mirrored in Study 3, in which strategically designed intervention texts were presented during the decision-making tasks to test their capacity to modify the participants’ inherent risk preference between the single and repeated gamble conditions and evaluate the persuasive efficacy of LLM-generated strategies on human decision biases.

Study 1 shows that the LLMs (GPT-3.5 and GPT-4) can successfully replicate the typical human pattern of risk aversion in single-play scenarios and risk seeking in repeated-play scenarios, though both models demonstrated an overall stronger tendency toward risk seeking compared with the human participants. Study 2 demonstrates that the human participants preferred low-EV certain options in single-play contexts and high-EV risky options in repeated-play contexts in both experiments. The participants also showed high agreement with the strategies generated by the LLMs in different scenarios. Study 3 confirms that the LLM-generated intervention texts can significantly influence the participants’ choice tendency in all four scenarios, with strong intervention effects observed in the single-play contexts. The LLM intervention strategies are characterised by reliance on expected value computations (normative) when promoting RCs and emphasis on certainty and robustness (descriptive) when promoting safe choices.

In summary, this study demonstrates that (1) LLMs can effectively simulate context-dependent human preferences in RC, particularly the shift from risk aversion in single plays to risk seeking in repeated plays; (2) LLMs can distinguish between the logic underlying single and repeated gambles and apply normative and descriptive reasoning accordingly to externalise decision strategies; and (3) the decision strategies extracted from LLM-generated reasoning can be used to construct effective intervention texts that can alter human preferences in classic risk decision tasks, thereby validating the feasibility and effectiveness of an LLM-based cognitive intervention pathway. This study offers a new technological paradigm for AI-assisted decision intervention and expands the application boundary of LLMs to human cognitive process modelling and regulation.

Music is the life of poetry. In classical Chinese poetry, the ping-ze (level-oblique) tonal pattern is a typical expression of poetic musicality. Using regulated five-character verse as materials, the present study manipulated both ping-ze patterns and semantic properties, and employed EEG to investigate the cognitive and neural mechanisms of ping-ze perception and poetic-meaning comprehension during reading of classical Chinese poetry. Behaviorally, both the main effect of ping-ze and the interaction between ping-ze and semantics were significant. ERP results further showed ping-ze effects during poetry reading, manifested as an early P200 effect, followed by an N400 effect and a late LPC effect. Moreover, ping-ze perception and poetic-meaning comprehension mutually regulated and integrated at mid-to-late stages, as reflected by interactions on the N400 and LPC. To reveal the relationship between neural activity and tonal/semantic features, we conducted decoding analyses on EEG data and found that neural networks trained on EEG signals could effectively classify different types of poetic lines. Together, the findings indicate that readers expect both the harmony of speech sounds themselves and the coordination between sound and meaning during poetry reading; such expectations influence early phonological representations and subsequent sound-meaning integration.

Temporal attention refers to the ability of individuals to prioritize information processing based on the timing of stimulus occurrence, which is critical for behavioral responses in daily life. However, whether rhythmic temporal attention is modulated by consciousness state remains unclear. In this study, we employed high-frequency flicker stimulation to manipulate the perceptual awareness level of visual rhythmic stimuli, integrating behavioral measures, hierarchical drift-diffusion model (HDDM) analysis, event-related potentials (ERP), and time-frequency analysis to systematically investigate the modulatory effects of consciousness state on rhythmic temporal attention, as well as the differential processing mechanisms of rhythmic cues at sub-second and supra-second timescales. Results from Experiment 1 showed that rhythmic cues elicited temporal attention effects under both conscious and unconscious states, but the effect was significantly attenuated in the unconscious condition. HDDM analysis further revealed that under conscious states, rhythmic cues reduced individuals’ decision boundaries, suggesting activation of endogenous processing at the decisional level, whereas this effect was absent in unconscious states. Building upon this, Experiment 2 found that the contingent negative variation (CNV) component and alpha oscillation suppression were more pronounced under conscious conditions, further supporting that consciousness state enhance temporal attention by modulating cognitive preparation and attention maintenance mechanisms. Moreover, although the inter-stimulus interval (ISI) of rhythmic cues did not affect the magnitude of temporal attention effects, overall responses were faster in supra-second interval conditions, consistent with predictions from the range-synthetic model of temporal cognition. Taken together, these findings suggest that rhythmic temporal attention is not only dependent on external rhythmic entrainment but may also involve endogenous decision-making mechanisms modulated by consciousness levels.

Theta oscillations are closely associated with cognitive control and are known to be involved in processing cross-modal stimulus conflict and response conflict. However, the relationship between these oscillations and the degree of such conflicts remains unclear. To address this issue, we used an audiovisual Stroop task with a 2-to-1 stimulus-response mapping and a time-resolved behavioral approach. The results demonstrated that processing of the task-relevant stimuli was rhythmically modulated by the task-irrelevant stimuli. When the task-irrelevant stimuli were the same as or different from the task-relevant stimuli, response times exhibited rhythmic fluctuations in the theta band. Furthermore, our results demonstrate that cognitive control processes exhibit theta oscillations at the behavioral level and that these oscillations directly modulate the magnitude of cross-modal stimulus conflict and response conflict. These findings elucidate the relationship between theta oscillations and conflict magnitude and extend the rhythmic theory of attention to the domain of cognitive control in conflict processing.

The present study integrated self-compassion with implementation intention, constructed a standardized implementation intention-based self-compassion (II-SC) sentence collection (Experiment 1) through 39 typical negative scenarios encountered by college students in daily life, and evaluated its efficacy in emotion regulation for students with depressive tendencies (Experiments 2-3). Results indicated: (1) The 117 II-SC sentences effectively reflected three self-compassion goals (self-kindness, common humanity, mindfulness) and significantly reduced emotional impact in negative situations (decreased arousal, increased valence), with the sentence collection demonstrated robust internal consistency and test-retest reliability, (2) When cognitive resources were abundant, II-SC sentences showed comparable emotion regulation effects to traditional self-compassion (TSC) sentences but required less cognitive effort, (3) When cognitive resources were scarce, II-SC demonstrated increased emotion regulation efficacy compared to TSC without additional cognitive cost, with no significant differences in physiological regulation effects on emotion-related indicators (heart rate, skin conductance). These findings suggest that the II-SC sentence collection facilitates automated and efficient emotion regulation processes for college students with depressive tendencies, enhances the effectiveness of self-compassion strategies, and provides potential solutions for coping with diverse stressors in daily life.

Third-party intervention plays a critical role in maintaining large-scale human group cooperation. Previous studies on children’s third-party intervention mainly focused on explicit norm violations, such as unfair distribution or damage to others’ belongings. Different from explicit violations, free-riding in group cooperation, where an individual benefits from others’ contributions without incurring the corresponding costs) involves implicit norms and lacks clearly identifiable victims. Recognizing and addressing free riders in group cooperation from a third-party perspective is an important aspect of children’s social norms acquisition and the development of cooperative behavior. This study examined the development and motivations of third-party intervention preference among children aged 4 to 11 years old in scenarios involving group cooperation norms violations.

In Study 1, 141 children (70 boys, Age range: 4.06~11.96 years, Mage = 8.02, SD = 2.30) heard a cooperation story in which one group member shirked effort by playing football instead of helping clean the classroom but still shared in the reward. Children’s moral evaluation, anger toward free-riding behavior, and their intervention preferences were measured in sequence. In Study 2, 125 children (63 boys, Age range: 4.34~11.68 years, Mage = 8.01, SD = 2.31) completed a similar task involving material contribution to a collective resource. To test cross-situational stability and rule out majority-influence effects, the group composition was adjusted to two cooperators and two free riders. In this story, a four-member group jointly dropped gold coins into a magic jar to get more gold coins. Among them, two members of the group each dropped one gold coin into the magic jar, while the other two members did not. In the end, the two coins that were dropped into the magic jar turned into four gold coins, and then each member of the group received one gold coin. Similarly to Study 1, children’s moral evaluation, anger responses, intervention preferences and motivations were measured.

Across both studies, children consistently evaluated free-riding negatively and reported anger toward it. As third-parties, children preferred intervention in free-riding behavior over non-intervention across all ages. With age, children’s preferences for third-party intervention showed a cross-situational stable developmental trend, gradually shifting from rewarding cooperators to punishing free riders. Importantly, the shift occurred earlier in labor non-contribution context (ages 5.51~5.67) than in material non-contribution context (ages 8.21~8.22). In terms of the motivations of intervention preferences, Study 2 found that the internal motivations for children to reward cooperators reflect both deontological motivation and consequentialist motivation. In contrast, the motivation to punish free-riders changes with age, gradually shifting toward the consequentialist motivation after age six. These findings indicated that the underlying motivations driving children’s intervention preferences are both specific and age-dependent.

These findings demonstrate that even young children can morally evaluate norm violations in group cooperation and engage in third-party interventions. Their intervention preferences develop in a stable, cross-situational manner, gradually shifting from rewarding prosocial behavior to punishing transgressors. However, the motivations underlying reward and punishment are distinct and age-dependent. This research provides valuable insights into the development of children’s cooperative behavior and their understanding of group cooperation norm.

Volunteering is a typical form of prosocial behavior in which individuals devote their time, energy, and skills to benefit others and society. Volunteering plays a vital role in enhancing volunteers’ individual well-being and mental health, strengthening community governance, promoting social cohesion, fostering economic development, and advancing societal progress. Despite these well-documented benefits, China’s volunteering sector continues to face several practical challenges, including low public willingness to participate, limited sustainability of engagement, and high volunteer turnover rates. Thus, there is a crucial need to identify factors that can promote long-term and sustainable volunteer participation. Studies to date have primarily focused on enhancing volunteers’ motivation and role identity. In contrast, the role of habit formation in sustaining volunteering has received relatively little scholarly attention. Moreover, the relationships linking intrinsic and extrinsic rewards, context stability, and volunteering habits remain insufficiently understood. To address these gaps, three studies were conducted to systematically examine how intrinsic and extrinsic rewards and context stability influence volunteering habits.

In Study 1, data were collected using semi-structured interviews with 25 volunteers who exhibited strong volunteering habits. Thematic analysis identified three types of intrinsic rewards (meaningfulness, happiness, and self-worth) and three types of extrinsic rewards (honorary rewards, material rewards, and social support and recognition). In addition, five forms of context stability (time, place, activity type, people, and mood) were identified as well. These factors may be potential predictors of volunteering habits.

Study 2 built on these qualitative findings through the development of quantitative measures of intrinsic rewards, extrinsic rewards, and context stability. A total of 1, 572 community volunteers and 853 student volunteers were surveyed using these scales and the Self-Report Habit Index for Volunteering. Multiple linear regression analyses revealed that intrinsic rewards, extrinsic rewards, and context stability significantly and positively predicted the strength of volunteering habits. Furthermore, a significant interaction between intrinsic rewards and context stability was found, such that when context stability was low, the positive effect of intrinsic rewards on volunteering habits became stronger. In contrast, the interaction between extrinsic rewards and context stability was not significant. These patterns were consistent across both community and student volunteer samples.

To complement the cross-sectional nature of Study 2, Study 3 adopted a longitudinal design in which three waves of data were collected, separated by three-month intervals, from 623 volunteers. Cross-lagged panel modeling revealed that intrinsic rewards at Time 1 (T1) significantly predicted volunteering habits at Time 2 (T2), whereas extrinsic rewards at T1 only marginally predicted volunteering habits at T2, and context stability at T1 did not significantly predict volunteering habits at T2. Similarly, intrinsic rewards at T2 significantly predicted volunteering habits at Time 3 (T3), whereas extrinsic rewards and context stability at T2 did not exhibit significant predictive effects on volunteering habits at T3. Furthermore, the interaction between intrinsic rewards and context stability at T2 was significant, while the interaction between extrinsic rewards and context stability at T2 was only marginally significant. Specifically, when context stability at T2 was low, the positive effects of either intrinsic or extrinsic rewards at T2 on volunteering habits at T3 became stronger.

In conclusion, intrinsic rewards (meaningfulness, happiness, and self-worth), extrinsic rewards (honorary rewards, material rewards, and social support and recognition), and context stability (time, place, activity type, people, and mood stabilities) emerged as key factors that enhance the strength of volunteering habits. Of these factors, intrinsic rewards were the most stable and robust predictors. Moreover, both intrinsic and extrinsic rewards interacted with context stability to predict volunteering habits, indicating that rewards exert stronger effects when the volunteering context is less stable. These findings offer valuable practical implications for sustaining long-term volunteer engagement. For temporary or emergency volunteer activities (e.g., natural disaster relief and large-scale social events) when service time and location are unstable and consistent contextual cues are lacking, it becomes particularly important to provide diverse forms of intrinsic and extrinsic rewards (such as honorary recognition, material subsidies, or social acknowledgment). Such incentives can enhance volunteers’ willingness to participate, encourage their repeated engagement, and ultimately facilitate the formation of volunteering habits. At the theoretical level, these results extend habit theory to the volunteering domain and offer new insights into how reward structures and contextual features can jointly facilitate the maintenance of prosocial behaviors.

In psychological research, determining the dimensional structure and characteristics of psychological traits is of paramount importance. Exploratory Factor Analysis (EFA) serves as a critical statistical methodology for identifying latent dimensions. Accurately determining the number of factors constitutes a pivotal technical challenge in EFA; under- or over-extraction of factors invariably yields detrimental consequences. To address this challenge, the present study conceptualizes eigenvalues as sequential data and employs a deep neural network architecture based on Long Short-Term Memory (LSTM) networks. Comprehensive evaluation metrics (including accuracy, precision, recall, F1-score, and Kappa) all exceeded 83%. Rigorous validation through extensive simulation studies and empirical analyses confirmed the robust performance of LSTM across diverse data conditions. Results demonstrate that LSTM achieves substantially higher accuracy than Comparison Data Fit (CDF), Empirical Kaiser Criterion (EKC), and Parallel Analysis (PA) methods, with a mean improvement rate of 48.50% and a peak improvement of 171.09%. Furthermore, LSTM exhibits smaller bias and superior robustness relative to CDF, EKC, and PA. Researchers may utilize the R package LSTMfactors to apply the LSTM model trained in this study to empirical data analysis.