Large language models (LLMs) are increasingly used to simulate human behavior in social settings such as legal mediation, negotiation, and dispute resolution. However, it remains unclear whether these simulations reproduce the personality–behavior patterns observed in humans. Human personality, for instance, shapes how individuals navigate social interactions, including strategic choices and behaviors in emotionally charged interactions. This raises the question: Can LLMs, when prompted with personality traits, reproduce personality-driven differences in human conflict behavior? To explore this, we introduce an evaluation framework that enables direct comparison of human-human and LLM-LLM behaviors in dispute resolution dialogues with respect to Big Five Inventory (BFI) personality traits. This framework provides a set of interpretable metrics related to strategic behavior and conflict outcomes. We additionally contribute a novel dataset creation methodology for LLM dispute resolution dialogues with matched scenarios and personality traits with respect to human conversations. Finally, we demonstrate the use of our evaluation framework with three contemporary closed-source LLMs and show significant divergences in how personality manifests in conflict across different LLMs compared to human data, challenging the assumption that personality-prompted agents can serve as reliable behavioral proxies in socially impactful applications. Our work highlights the need for psychological grounding and validation in AI simulations before real-world use.

Negotiation requires dynamically balancing self-interest and cooperation within the flow of conversation to maximize one's own utility.Yet, existing agents struggle due to bounded rationality in human data, low adaptability to counterpart behavior, and limited strategic reasoning.To address this, we introduce principle-driven negotiation agents, powered by ASTRA, a novel framework for turn-level offer optimization grounded in two core principles: opponent modeling and Tit-for-Tat reciprocity.ASTRA operates in three stages: (1) interpreting counterpart behavior, (2) optimizing counteroffers via a tool-integrated action with a linear programming (LP) solver, and (3) selecting offers based on strategy assessment and the partner's acceptance probability.Through simulations and human evaluations, our agent effectively adapts to an opponent's shifting stance and achieves favorable outcomes through enhanced adaptability and strategic reasoning.Beyond enhancing negotiation performance, it also serves as a powerful coaching tool, offering interpretable strategic feedback and optimal offer recommendations beyond human bounded rationality, with its potential further validated through human evaluation.

Large Language Models (LLMs) are increasingly deployed in socially complex, interactiondriven tasks, yet their ability to mirror human behavior in emotionally and strategically complex contexts remains underexplored.This study assesses the behavioral alignment of personality-prompted LLMs in adversarial dispute resolution by simulating multi-turn conflict dialogues that incorporate negotiation.Each LLM is guided by a matched Five-Factor personality profile to control for individual variation and enhance realism.We evaluate alignment across three dimensions: linguistic style, emotional expression (e.g., anger dynamics), and strategic behavior.GPT-4.1 achieves the closest alignment with humans in linguistic style and emotional dynamics, while Claude-3.7-Sonnetbest reflects strategic behavior.Nonetheless, substantial alignment gaps persist.Our findings establish a benchmark for alignment between LLMs and humans in socially complex interactions, underscoring both the promise and the limitations of personality conditioning in dialogue modeling.Avg. Within-Human JSD: 0.179 Avg.

Language-driven generative agents have enabled large-scale social simulations with transformative uses, from interpersonal training to aiding global policy-making.However, recent studies indicate that generative agent behaviors often deviate from expert expectations and real-world data-a phenomenon we term the Behavior-Realism Gap.To address this, we introduce a theoretical framework called Persona-Environment Behavioral Alignment (PEBA), formulated as a distribution matching problem grounded in Lewin's behavior equation stating that behavior is a function of the person and their environment.Leveraging PEBA, we propose PersonaEvolve (PEvo), an LLM-based optimization algorithm that iteratively refines agent personas, implicitly aligning their collective behaviors with realistic expert benchmarks within a specified environmental context.We validate PEvo in an active shooter incident simulation we developed, achieving an 84% average reduction in distributional divergence compared to no steering and a 34% improvement over explicit instruction baselines.Results also show PEvo-refined personas generalize to novel, related simulation scenarios.Our method greatly enhances behavioral realism and reliability in high-stakes social simulations.More broadly, the PEBA-PEvo framework provides a principled approach to developing trustworthy LLM-driven social simulations.1

Artificial intelligence (AI) decision support systems are crucial in modern decision-making processes. Their increasing human-like adaptability introduces challenges, especially when their recommendations, for whatever reason, need to conflict with user preferences. This study examines the communication strategies AI systems should employ when their recommendations conflict with user preferences. We explored this research question through a hypothetical future interface where ChatGPT offers travel recommendations populated on a map. An online survey-based experiment was conducted, presenting 160 participants with ChatGPT-generated travel recommendations displayed alongside Bing map visuals. We employed a mixed-method experimental design, combining both between-subjects and within-subjects approaches, to investigate the impact of conflicting recommendations and the acknowledgment of user preferences on the acceptance of these recommendations. This effect is especially pronounced when the AI system acknowledges users’ preferences yet still offers conflicting recommendations to them. Contrary to the expectation that acknowledging users’ preferences could buffer the impact of such conflicts, our observations indicate the contrary. The presence of conflict following acknowledgment of users’ preferences, significantly causes a backfire effect, leading users to reject the recommendations. These findings underscore the need for consideration of recommendation delivery strategies in AI decision support systems and offer insights for designing future user interfaces and user experience research in the realm of recommendations provided by AI decision-support systems.

Social media platforms provide various privacy settings, which users can adjust to fit their privacy needs. Platforms claim that this is sufficient – users have power to accept the default settings they like, and change those they do not like. In this paper, we seek to quantify user awareness of, preferences around and ability to adjust social media privacy settings. We conduct an online survey of 541 participants across six different social media platforms: Facebook, Instagram, X, LinkedIn, TikTok, and Snapchat. We focus on nine privacy settings that are commonly available across these platforms, and evaluate participants’ preferences for privacy, awareness of the privacy settings and ability to locate them. We find that default settings are ill-aligned with user preferences – 92% of participants prefer at least one of the privacy options to be more private than the default. We further find that users are generally not aware of privacy settings, and struggle to find them. 80% of participants have never seen at least one privacy setting, and 79% of participants rated at least one setting as hard to find. We also find that the fewer privacy settings a user has seen, the harder for them to locate those settings, and the higher the level of privacy they desire. Additionally, we find that there are significant differences in privacy setting preferences and usability across different user age groups and across platforms. Older users are more conservative about their privacy, they have seen significantly fewer privacy settings, and they spend significantly more time locating them than younger users. On some platforms, like LinkedIn, users opt for higher visibility, while on others they prefer more privacy. Some platforms, like TikTok, make it significantly easier for users to locate privacy settings. Based on our findings, we provide recommendations on default values and how to improve usability of privacy settings on social media.

The use of social robots in service tasks is spreading, showcasing advantages for both consumers and service providers. However, their widespread adoption is hindered by a notable lack of trust. Our study aims to uncover insights into the factors influencing the adoption of social robots in home settings, exploring the factors that lead users to trust and eventually adopt robots. We designed two experimental conditions, presenting the Amazon Astro robot from different perspectives (high-angle and eye-level) and demonstrating its different abilities to 198 people recruited from MTurk. We employed both quantitative (trust, first impressions of warmth and competence as well as usability, familiarity, and attitudes) questionnaires and qualitative (word analysis) assessments, and results showed that participants had higher trust scores when seeing the robot from an eye-level perspective. In addition, usability, familiarity and competence were shown to explain a significant amount of variance in trust. While existing negative attitudes towards robots and the participants’ age were shown to be the strongest predictors for participants’ willingness to purchase a robot, trust was able to significantly affect use intention. We contribute to the broader understanding of the challenges and opportunities in integrating social robots into daily life, shedding light on the dynamics between technological innovation and consumer adoption.

Previous studies have solely focused on establishing Machine Learning (ML) models for automated detection of stress arousal. However, these studies do not recognize stress appraisal and presume stress is a negative mental state. Yet, stress can be classified according to its influence on individuals; the way people perceive a stressor determines whether the stress reaction is considered as eustress (positive stress) or distress (negative stress). Thus, this study aims to assess the potential of using an ML approach to determine stress appraisal and identify eustress and distress instances using physiological and behavioral features. The results indicate that distress leads to higher perceived stress arousal compared to eustress. An XGBoost model that combined physiological and behavioral features using a 30 second time window had 83.38% and 78.79% F1-scores for predicting eustress and distress, respectively. Gender-based models resulted in an average increase of 2-4% in eustress and distress prediction accuracy. Finally, a model to predict the simultaneous assessment of eustress and distress, distinguishing between pure eustress, pure distress, eustress-distress coexistence, and the absence of stress achieved a moderate F1-score of 65.12%. The results of this study lay the foundation for work management interventions to maximize eustress and minimize distress in the workplace.

Negotiation is a complex social interaction that encapsulates emotional encounters in human decision-making. Virtual agents that can negotiate with humans by the means of language are useful in pedagogy and conversational AI. To advance the development of such agents, we explore the role of emotion in the prediction of two important subjective goals in a negotiation – outcome satisfaction and partner perception. We devise ways to measure and compare different degrees of emotion expression in negotiation dialogues, consisting ofemoticon,lexical, andcontextualvariables. Through an extensive analysis of a large-scale dataset in chat-based negotiations, we find that incorporating emotion expression explains significantly more variance, above and beyond the demographics and personality traits of the participants. Further, our temporal analysis reveals that emotive information from both early and later stages of the negotiation contributes to this prediction, indicating the need for a continual learning model of capturing emotion for automated agents. Finally, we extend our analysis to another dataset, showing promise that our findings generalize to more complex scenarios. We conclude by discussing our insights, which will be helpful for designing adaptive negotiation agents that interact through realistic communication interfaces.

Within the field of Affective Computing, facial expressions have traditionally been used as a means of inferring valence and emotions. Most studies have focused on interpreting facial expressions as an isolated signal, typically training algorithms with annotators with a $3^{\text {rd }}$-person point of view, often without access to the original context. When the context is provided, recent research highlights that the interpretation annotators assign to facial expressions is sometimes more strongly influenced by the context than the facial expression. But even when the context is provided, annotators are psychologically and physically distant from the original setting that produced the emotion. In this paper, we explore how context and facial expressions shape $2^{\text {nd }}$-person interpretation of facial expressions and compare this to $1^{\text {st }}$-person self-report. Results show that both expression and context contribute to self and other impressions but in different ways. Expressions and context are independent predictors of $1^{\text {st }}$-person judgments but interact to determine $2^{\text {nd }}$. person judgments. In particular, the way players interpret their partner’s facial cues changes dramatically based on what just occurred in the game. We discuss the implication of these findings for automatic emotion recognition methods.

Dialogue systems capable of social influence such as persuasion, negotiation, and therapy, are essential for extending the use of technology to numerous realistic scenarios.However, existing research primarily focuses on either task-oriented or open-domain scenarios, a categorization that has been inadequate for capturing influence skills systematically.There exists no formal definition or category for dialogue systems with these skills and data-driven efforts in this direction are highly limited.In this work, we formally define and introduce the category of social influence dialogue systems that influence users' cognitive and emotional responses, leading to changes in thoughts, opinions, and behaviors through natural conversations.We present a survey of various tasks, datasets, and methods, compiling the progress across seven diverse domains.We discuss the commonalities and differences between the examined systems, identify limitations, and recommend future directions.This study serves as a comprehensive reference for social influence dialogue systems to inspire more dedicated research and discussion in this emerging area.

A natural way to design a negotiation dialogue system is via self-play RL: train an agent that learns to maximize its performance by interacting with a simulated user that has been designed to imitate human-human dialogue data.Although this procedure has been adopted in prior work, we find that it results in a fundamentally flawed system that fails to learn the value of compromise in a negotiation, which can often lead to no agreements (i.e., the partner walking away without a deal), ultimately hurting the model's overall performance.We investigate this observation in the context of DealOrNoDeal task, a multi-issue negotiation over books, hats, and balls.Grounded in negotiation theory from Economics, we modify the training procedure in two novel ways to design agents with diverse personalities and analyze their performance with human partners.We find that although both techniques show promise, a selfish agent, which maximizes its own performance while also avoiding walkaways, performs superior to other variants by implicitly learning to generate value for both itself and the negotiation partner.We discuss the implications of our findings for what it means to be a successful negotiation dialogue system and how these systems should be designed in the future.

Negotiation is a complex social interaction that encapsulates emotional encounters in human decision-making. Virtual agents that can negotiate with humans are useful in pedagogy and conversational AI. To advance the development of such agents, we explore the prediction of two important subjective goals in a negotiation – outcome satisfaction and partner perception. Specifically, we analyze the extent to which emotion attributes extracted from the negotiation help in the prediction, above and beyond the individual difference variables. We focus on a recent dataset in chat-based negotiations, grounded in a realistic camping scenario. We study three degrees of emotion dimensions – emoticons, lexical, and contextual by leveraging affective lexicons and a state-of-the-art deep learning architecture. Our insights will be helpful in designing adaptive negotiation agents that interact through realistic communication interfaces.

Women earn less than men in technical fields. Competing theories have been offered to explain this disparity. Some argue that women underperform in negotiating their salary, in-part due to language in job descriptions, called gender triggers, which leave women feeling disadvantaged in salary negotiations. Others point to structural and institutional bias: i.e., recruiters make better offers to men even when women exhibit equal negotiation skills. As a final salary is co-constructed though an interaction between employees and recruiters, it is difficult to disentangle these views. Here, we discuss how intelligent virtual agents serve as powerful methodological tools that lend new insight into this psychological debate. We use virtual negotiators to examine the impact of gender triggers on computer science (CS) undergraduates that engaged in a simulated salary negotiation with an automated recruiter. We find that, regardless of gender, CS students are reluctant to negotiate, and this hesitancy likely lowers their starting salary. Even when they negotiate, students show little skill in discovering tradeoffs that could enhance their salary, highlighting the need for negotiation training in technical fields. Most importantly, we find little evidence that gender triggers impact women's negotiated outcomes, at least within the field of CS. We argue that findings that emphasize women's individual deficits may reflect a lack of experimental control, which intelligent agents can help correct, and that structural and institutional explanations of inequity deserve greater attention.

Kushal Chawla, Jaysa Ramirez, Rene Clever, Gale Lucas, Jonathan May, Jonathan Gratch. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Heating, ventilation, and air conditioning (HVAC) systems account for 43% of building energy consumption, yet only 38% of commercial building occupants are satisfied with the thermal environment. The primary reasons for low occupant satisfaction are that HVAC operations do not integrate occupant comfort requirements nor control the thermal environment at the individual level. Personal comfort systems (PCSs) enable local control of the thermal environment around each occupant. However, full manual control of PCS can be inefficient, and fully automated PCS reduces an occupant's perceived control over the environment, which can then lead to lower satisfaction. A better solution might lie somewhere between fully manual and fully automated environmental control. In this article, we describe the development and implementation of an Internet-of-Things (IoT)-based intelligent agent that learns individual occupant comfort requirements and controls the thermal environment using PCS (i.e., a local fan and a heater). We tested different levels of automation where control is shared between an intelligent agent and the end user. Our results show that PCS use improves occupant satisfaction and including some level of automation can improve occupant satisfaction further than what is possible with manually operated PCS. Among the levels of automation investigated, inquisitive automation, where the user approves/declines the control actions of the intelligent agent before execution, led to highest occupant satisfaction with the thermal environment.

Agents that interact with humans in complex, social tasks need the ability to comprehend as well as employ common social strategies. In negotiation, there is ample evidence of such techniques being used efficaciously in human interchanges. In this work, we demonstrate a new design for socially aware agents that employ one such technique---favor exchange---in order to gain value when playing against humans. In an online study of a robust, simulated social negotiation task, we show that these agents are effective against real human participants. In particular, we show that agents that ask for favors during the course of a repeated set of negotiations are more successful than those that do not. Additionally, previous work has demonstrated that humans can detect when agents betray them by failing to return favors that were previously promised. By contrast, this work indicates that these betrayal techniques may go largely undetected in complex scenarios.

Research has shown that virtual agents can be effective tools for teaching negotiation. Virtual agents provide an opportuni-ty for students to practice their negotiation skills which leads to better outcomes. However, these negotiation training agents often lack the ability to understand the errors students make when negotiating, thus limiting their effectiveness as training tools. In this article, we argue that automated opponent-modeling techniques serve as effective methods for diagnos-ing important negotiation mistakes. To demonstrate this, we analyze a large number of participant traces generated while negotiating with a set of automated opponents. We show that negotiators' performance is closely tied to their understanding of an opponent's preferences. We further show that opponent modeling techniques can diagnose specific errors includ-ing: failure to elicit diagnostic information from an opponent, failure to utilize the information that was elicited, and failure to understand the transparency of an opponent. These results show that opponent modeling techniques can be effective methods for diagnosing and potentially correcting crucial ne-gotiation errors.

Recent research shows that how we respond to other social actors depends on what sort of mind we ascribe to them. In this article we examine how perceptions of a virtual agent's mind shape behavior in human-agent negotiations. We varied descriptions and communicative behavior of virtual agents on two dimensions according to the mind perception theory:agency (cognitive aptitude) andpatiency (affective aptitude). Participants then engaged in negotiations with the different agents. People scored more points and engaged in shorter negotiations with agents described to be cognitively intelligent, and got lower points and had longer negotiations with agents that were described to be cognitively unintelligent. Accordingly, agents described as having low agency ended up earning more points than those with high agency. Within the negotiations themselves, participants sent more happy and surprise emojis and emotionally valenced messages to agents described to be emotional. This high degree of described patiency also affected perceptions of the agent's moral standing and relatability. In short, manipulating the perceived mind of agents affects how people negotiate with them. We discuss these results, which show that agents are perceived not only as social actors, but as intentional actors through negotiations.

Research in social science distinguishes between two types of social influence: informational and normative. Informational social influence is driven by the desire to evaluate ambiguous situations correctly, whereas normative social influence is driven by the desire to be liked and gain social acceptance from another person. Although we know from research that humans can effectively use either of these techniques to persuade other humans, scholars have yet to examine the relative effectiveness of informational versus normative social influence when used by virtual agents. We report a study in which users interact with a system that persuades them either using informational or normative social influence. Furthermore, to compare agents to human interlocutors, users are told that the system is either tele-operated by a human (avatar) or fully-automated (agent). Using this design, we are able to compare the effectiveness of virtual agents (vs humans) in employing informational versus normative social influence. Participants interacted with the system, which employed a Wizard-of-Oz operated virtual agent that tried to persuade the user to agree with its rankings on a "survival task." Controlling for initial divergence in rankings between user and the agent, there was a significant main effect such that informational social influence resulted in greater influence than normative influence. However, this was qualified by an interaction that approached significance; users were, if anything, more persuaded by informational influence when they believe the agent was AI (compared to a human), whereas there was no difference between the agent and avatar in the normative influence condition.

Socially intelligent artificial agents and robots are anticipated to become ubiquitous in home, work, and military environments. With the addition of such agents to human teams it is crucial to evaluate their role in the planning, decision making, and conflict mediation processes. We conducted a study to evaluate the utility of a virtual agent that provided mission planning support in a three-person human team during a military strategic mission planning scenario. The team consisted of a human team lead who made the final decisions and three supporting roles, two humans and the artificial agent. The mission outcome was experimentally designed to fail and introduced a conflict between the human team members and the leader. This conflict was mediated by the artificial agent during the debriefing process through discuss or debate and open communication strategies of conflict resolution [1]. Our results showed that our teams experienced conflict. The teams also responded socially to the virtual agent, although they did not find the agent beneficial to the mediation process. Finally, teams collaborated well together and perceived task proficiency increased for team leaders. Socially intelligent agents show potential for conflict mediation, but need careful design and implementation to improve team processes and collaboration.

Behavioral intervention strategies have yet to become successful in the development of initiatives to foster pro-environmental behaviors in buildings. In this paper, we explored the potentials of increasing the effectiveness of requests aiming to promote pro-environmental behaviors by engaging users in a social dialog, given the effects of two possible personas that are more related to the buildings (i.e., building vs. building manager). We tested our hypotheses and evaluated our findings in virtual and physical environments and found similar effects in both environments. Our results showed that social dialog involvement persuaded respondents to perform more pro-environmental actions. However, these effects were significant when the requests were delivered by an agent representing the building. In addition, these strategies were not equally effective across all types of people and their effects varied for people with different characteristics. Our findings provide useful design choices for persuasive technologies aiming to promote pro-environmental behaviors.

A crucial aspect of virtual gaming experiences is the avatar: the player's virtual self-representation. While research has demonstrated benefits to using self-similar avatars in some virtual experiences, such avatars sometimes produce a more negative experience for women. To help researchers and game designers assess the cost-benefit tradeoffs of self-similar avatars, we compared players' performance and subjective experience in a search and rescue computer game when using two different photorealistic avatars: their own self or a friend, and when playing either a social (rescuing people) or a nonsocial (rescuing gems) version of the game. There was no effect of avatar appearance on players' performance or subjective experience in either game version, but we also found that women's experience with self-similar avatars was no more negative than men's. Our results suggest that avatar appearance may not make a difference to players in certain game contexts.

This work explores the extent to which social dialogue can mitigate (or exacerbate) the loss of trust caused when robots make conversational errors. Our study uses a NAO robot programmed to persuade users to agree with its rankings on two tasks. We perform two manipulations: (1) The timing of conversational errors - the robot exhibited errors either in the first task, the second task, or neither; (2) The presence of social dialogue - between the two tasks, users either engaged in a social dialogue with the robot or completed a control task. We found that the timing of the errors matters: replicating previous research, conversational errors reduce the robot's influence in the second task, but not on the first task. Social dialogue interacts with the timing of errors, acting as an intensifier: social dialogue helps the robot recover from prior errors, and actually boosts subsequent influence; but social dialogue backfires if it is followed by errors, because it extends the period of good performance, creating a stronger contrast effect with the subsequent errors. The design of social robots should therefore be more careful to avoid errors after periods of good performance than early on in a dialogue.

A long tradition of research suggests a relationship between emotional mimicry and pro-social behavior, but the nature of this relationship is unclear. Does mimicry cause rapport and cooperation, or merely reflect it? Virtual humans can provide unique insights into these social processes by allowing unprecedented levels of experimental control. In a 2 x 2 factorial design, we examined the impact of facial mimicry and counter-mimicry in the iterated prisoner's dilemma. Participants played with an agent that copied their smiles and frowns or one that showed the opposite pattern -- i.e., that frowned when they smiled. As people tend to smile more than frown, we independently manipulated the contingency of expressions to ensure any effects are due to mimicry alone, and not the overall positivity/negativity of the agent: i.e., participants saw either a reflection of their own expressions or saw the expressions shown to a previous participant. Results show that participants smiled significantly more when playing an agent that mimicked them. Results also show a complex association between smiling, feelings of rapport, and cooperation. We discuss the implications of these findings on virtual human systems and theories of cooperation.

This work explores whether culture impacts the extent to which social dialogue can mitigate (or exacerbate) the loss of trust caused when agents make conversational errors. Our study uses an agent designed to persuade users to agree with its rankings on two tasks. Participants from the U.S. and Japan completed our study. We perform two manipulations: (1) The presence of conversational errors -- the agent exhibited errors in the second task or not; (2) The presence of social dialogue -- between the two tasks, users either engaged in a social dialogue with the agent or completed a control task. Replicating previous research, conversational errors reduce the agent's influence. However, we found that culture matters: there was a marginally significant three-way interaction with culture, presence of social dialogue, and presence of errors. The pattern of results suggests that, for American participants, social dialogue backfired if it is followed by errors, presumably because it extends the period of good performance, creating a stronger contrast effect with the subsequent errors. However, for Japanese participants, social dialogue if anything mitigates the detrimental effect of errors; the negative effect of errors is only seen in the absence of a social dialogue. Agent design should therefore take the culture of the intended users into consideration when considering use of social dialogue to bolster agents against conversational errors.

Research has demonstrated promising benefits of applying virtual trainers to promote physical fitness. The current study investigated the value of virtual agents in the context of personal fitness, compared to trainers with greater levels of perceived agency (avatar or live human). We also explored the possibility that the effectiveness of the virtual trainer might depend on the affective tone it uses when trying to motivate users. Accordingly, participants received either positively or negatively valenced motivational messages from a virtual human they believed to be either an agent or an avatar, or they received the messages from a human instructor via skype. Both self-report and physiological data were collected. Like in-person coaches, the live human trainer who used negatively valenced messages were well-regarded; however, when the agent or avatar used negatively valenced messages, participants responded more poorly than when they used positively valenced ones. Perceived agency also affected rapport: compared to the agent, users felt more rapport with the live human trainer or the avatar. Regardless of trainer type, they also felt more rapport - and said they put in more effort - with trainers that used positively valenced messages than those that used negatively valenced ones. However, in reality, they put in more physical effort (as measured by heart rate) when trainers employed the more negatively valenced affective tone. We discuss implications for human--computer interaction.

We present the results of a study in which humans negotiate with computerized agents employing varied tactics over a repeated number of economic ultimatum games. We report that certain agents are highly effective against particular classes of humans: several individual difference measures for the human participant are shown to be critical in determining which agents will be successful. Asking for favors works when playing with pro-social people but backfires with more selfish individuals. Further, making poor offers invites punishment from Machiavellian individuals. These factors may be learned once and applied over repeated negotiations, which means user modeling techniques that can detect these differences accurately will be more successful than those that don't. Our work additionally shows that a significant benefit of cooperation is also present in repeated games---after sufficient interaction. These results have deep significance to agent designers who wish to design agents that are effective in negotiating with a broad swath of real human opponents. Furthermore, it demonstrates the effectiveness of techniques which can reason about negotiation over time.

Emotions have a temporal nature and very often personality traits and underlying psychological conditions are hidden in the dynamics of those expressions. Within this work, we investigate the dynamics of the facial displays of dyads during an iterated social dilemma. We focus on the effect of gender and gender-pairing on those behaviors. We use networks to capture the temporal dynamics and create measures of inter- and intra-personal dependencies of emotional states. Our analysis on an iterated prisoner's dilemma corpus suggests that there are gender differences on the transitions of the emotional states and the degree of emotional influence from the opponent.

Social robots establish rapport with human users. This work explores the extent to which rapport-building can benefit (or harm) conversations with robots, and under what circumstances this occurs. For example, previous work has shown that agents that make conversational errors are less capable of influencing people than agents that do not make errors [1]. Some work has shown this effect with robots, but prior research has not considered additional factors such as the level of rapport between the person and the robot. We predicted that building rapport through a social dialogue (such as an ice-breaker) could mitigate the detrimental effect of a robot's errors on influence. Our study used a Nao robot programmed to persuade users to agree with its rankings on two "survival tasks" (e.g., lunar survival task). We manipulated both errors and social dialogue:the robot either exhibited errors in the second survival task or not, and users either engaged in an ice-breaker with the robot between the two survival tasks or completed a control task. Replicating previous research, errors tended to reduce the robot's influence in the second survival task. Contrary to our prediction, results revealed that the ice-breaker did not mitigate the effect of errors, and if anything, errors were more harmful after the ice-breaker (intended to build rapport) than in the control condition. This backfiring of attempted rapport-building may be due to a contrast effect, suggesting that the design of social robots should avoid introducing dialogues of incongruent quality.

This paper builds on prior psychological studies that identify signals of trustworthiness between two human negotiators. Unlike prior work, the current work tracks such signals automatically and fuses them into computational models that predict trustworthiness. To achieve this goal, we apply automatic trackers to recordings of human dyads negotiating in a multi-issue bargaining task. We identify behavioral indicators in different modalities (facial expressions, gestures, gaze, and conversational features) that are predictive of trustworthiness. We predict both objective trustworthiness (i.e., are they honest) and perceived trustworthiness (i.e., do they seem honest to their interaction partner). Our experiments show that people are poor judges of objective trustworthiness (i.e., objective and perceived trustworthiness are predicted by different indicators), and that multimodal approaches better predict objective trustworthiness, whereas people overly rely on facial expressions when judging the honesty of their partner. Moreover, domain knowledge (from the literature and prior analysis of behaviors) facilitates the model development process.

Recent advances in scanning technology have enabled the widespread capture of 3D character models based on human subjects. Intuition suggests that, with these new capabilities to create avatars that look like their users, every player should have his or her own avatar to play video games or simulations. We explicitly test the impact of having one's own avatar (vs. a yoked control avatar) in a simulation (i.e., maze running task with mines). We test the impact of avatar identity on both subjective (e.g., feeling connected and engaged, liking avatar's appearance, feeling upset when avatar's injured, enjoying the game) and behavioral variables (e.g., time to complete task, speed, number of mines triggered, riskiness of maze path chosen). Results indicate that having an avatar that looks like the user improves their subjective experience, but there is no significant effect on how users perform in the simulation.

Reduced frequency range in vowel production is a well documented speech characteristic of individuals with psychological and neurological disorders. Affective disorders such as depression and post-traumatic stress disorder (PTSD) are known to influence motor control and in particular speech production. The assessment and documentation of reduced vowel space and reduced expressivity often either rely on subjective assessments or on analysis of speech under constrained laboratory conditions (e.g. sustained vowel production, reading tasks). These constraints render the analysis of such measures expensive and impractical. Within this work, we investigate an automatic unsupervised machine learning based approach to assess a speaker's vowel space. Our experiments are based on recordings of 253 individuals. Symptoms of depression and PTSD are assessed using standard self-assessment questionnaires and their cut-off scores. The experiments show a significantly reduced vowel space in subjects that scored positively on the questionnaires. We show the measure's statistical robustness against varying demographics of individuals and articulation rate. The reduced vowel space for subjects with symptoms of depression can be explained by the common condition of psychomotor retardation influencing articulation and motor control. These findings could potentially support treatment of affective disorders, like depression and PTSD in the future.

We present the SimSensei system, a fully automatic virtual agent that conducts interviews to assess indicators of psychological distress. We emphasize on the perception part of the system, a multimodal framework which captures and analyzes user state for both behavioral understanding and interactional purposes.

Sporting events evoke strong emotions amongst fans and thus act as natural laboratories to explore emotions and how they unfold in the wild. Computational tools, such as sentiment analysis, provide new ways to examine such dynamic emotional processes. In this article we use sentiment analysis to examine tweets posted during 2014 World Cup. Such analysis gives insight into how people respond to highly emotional events, and how these emotions are shaped by contextual factors, such as prior expectations, and how these emotions change as events unfold over time. Here we report on some preliminary analysis of a World Cup twitter corpus using sentiment analysis techniques. We show these tools can give new insights into existing theories of what makes a sporting match exciting. This analysis seems to suggest that, contrary to assumptions in sports economics, excitement relates to expressions of negative emotion. We also discuss some challenges that such data present for existing sentiment analysis techniques and discuss future analysis.

Even with the rise in use of TeleMedicine for health care and mental health, research suggests that clinicians may have difficulty reading nonverbal cues in computer-mediated situations. However, the recent progress in tracking affective markers (i.e., displays of emotional expressions on face and in voice) has opened the door to new clinical applications that might help health care providers better read nonverbal behaviors when employing TeleMedicine. For example, an interface that automatically quantified affective markers could assist clinicians in their assessment of and treatment for psychological distress (i.e., symptoms of depression and Post-traumatic Stress Disorder (PTSD)). To move towards this prospect, we will show that clinicians' judgments of these nonverbal affective markers (e.g., smile, frown, eye contact, tense voice) could be informed by such technology. The results of our evaluation suggest that clinicians' ratings of nonverbal affective markers are less predictive of psychological distress than automatically quantified affective markers. Because such quantifications are more strongly associated with psychological distress than clinician ratings of these same nonverbal behaviors, an affective interface providing quantifications of nonverbal affective markers could potentially improve assessment of psychological distress..

Negotiation between virtual agents and humans is a complex field that requires designers of systems to be aware not only of the efficient solutions to a given game, but also the mechanisms by which humans create value over multiple negotiations. One way of considering the agent's impact beyond a single negotiation session is by considering the use of external “ledgers” across multiple sessions. We present results that describe the effects of favor exchange on negotiation outcomes, fairness, and trust for two distinct cross-cultural populations, and illustrate the ramifications of their similarities and differences on virtual agent design.

Machiavellianism, by definition, is the tendency to use other people as a tool to achieve one's own goals. Despite the large focus on the Big Five traits of personality, this anti-social trait is relatively unexplored in the computational realm. Automatically recognizing anti-social traits can have important uses across a variety of applications. In this paper, we use negotiation as a setting that provides Machiavellians with the opportunity to reveal their exploitative inclinations. We use textual, visual, acoustic, and behavioral cues to automatically predict High vs. Low Machiavellian personalities. These learned models have good accuracy when compared with other personality-recognition methods, and we provide evidence that the automatically-learned models are consistent with existing literature on this anti-social trait, giving evidence that these results can generalize to other domains.

Emotional signaling plays an important role in negotiations and other social decision-making tasks as it can signal intention and shape joint decisions. Specifically it has been shown to influence cooperation or competition. This has been shown in previous studies for scripted interactions that control emotion signaling and rely on manual coding of affect. In this work we examine face-to-face interactions in an iterative social dilemma task (prisoner's dilemma) via an automatic framework for facial expression analysis. We explore if automatic analysis of emotion can give insight into the social function of emotion in face-to-face interactions. Our analysis suggests that positive and negative displays of emotion are associated with more prosocial and proself game acts respectively. Moreover signaling cooperative intentions to the opponent via positivity can leave participants more open to exploitation, whereas signaling a more tough stance via negativity seems to discourage exploitation. However, the benefit of negative affect is short-term and both players do worse over time if they show negative emotions.

We present the SimSensei system, a fully automatic virtual agent that conducts interviews to assess indicators of psychological distress. With this demo, we focus our attention on the perception part of the system, a multimodal framework which captures and analyzes user state behavior for both behavioral understanding and interactional purposes. We will demonstrate real-time user state sensing as a part of the SimSensei architecture and discuss how this technology enabled automatic analysis of behaviors related to psychological distress.

Recent advances in scanning technology have enabled the widespread capture of 3D character models based on human subjects. However, in order to generate a recognizable 3D avatar, the movement and behavior of the human subject should be captured and replicated as well. We present a method of generating a 3D model from a scan, as well as a method to incorporate a subjects style of gesturing into a 3D character. We present a study which shows that 3D characters that used the gestural style as their original human subjects were more recognizable as the original subject than those that don't.