Publications

Uncertainty estimation is essential for the trustworthy deployment of LLMs: it underpins self-refinement and serves as a decision signal for downstream systems, from verifiers to collaborating agents. However, existing approaches operate at suboptimal granularities: token-level scores lack semantic coherence, while sequence-level scores fail to localize errors. We formalize Span-Level Uncertainty Estimation (SLUE), a new task that targets the natural granularity for uncertainty: semantically coherent text spans, each conveying a single assessable unit of meaning. To address this task, we introduce SPANUQ, a lightweight probe that distills the uncertainty knowledge from expensive multi-sample inference into a single forward pass over LLM hidden states. SPANUQ employs a DETR-style span decoder to simultaneously detect spans and estimate their uncertainty via a Mixture of Beta distribution, trained with a principled combination of Beta NLL regression and contrastive ranking objectives. We construct SPANUQ-BENCH, the first span-level uncertainty benchmark comprising 20K prompts, 293K annotated spans, and continuous soft labels derived from multi-sample claim verification. Experiments on five LLM backbones show that SPANUQ consistently achieves the best span-level uncertainty quality , outperforming the strongest probe baseline and all sampling-based methods while being 10~20x faster. Its DETR-based span detector attains 0.910 F1, surpassing the best heuristic by 39.4%, enabling precise error localization that sequence-level methods cannot provide. The same architecture ports to five LLMs spanning two model families, with one probe trained per backbone, and we additionally observe that sequence-level uncertainty is partially decomposable, suggesting that span-level estimation subsumes sequence-level as a special case. The project page is available damon-demon.github.io/SpanUQ.

Reinforcement learning (RL) is effective in enhancing the accuracy of large language models in complex reasoning tasks. Existing RL policy optimization frameworks rely on final-answer correctness as feedback signals and rarely capture the internal logical structure of the reasoning process. Consequently, the models would generate fluent and semantically relevant responses but logically inconsistent, structurally erratic, or redundant. To this end, we propose StaRPO, a stability-augmented reinforcement learning framework that explicitly incorporates reasoning stability into the optimization objective. Our StaRPO decomposes stability into two computable lightweight metrics: the Autocorrelation Function (ACF) to evaluate local step-to-step coherence, and Path Efficiency (PE) to evaluate global goal-directedness of the reasoning trajectory. These stability rewards are combined with task rewards to provide complementary and process-aware feedback. We validate the effectiveness of using ACF and PE rewards by showing their correlation with logic errors on two backbone models. Experiments on four reasoning benchmarks show that StaRPO consistently outperforms compared baselines and can enhance both final-answer accuracy and logical stability.

With the growing availability of quantitative forecasts from various sources, effectively communicating these multiple forecasts has become increasingly crucial. Recent advances have explored using Multiple-Forecast Visualizations (MFVs) to display multiple time-series forecasts. However, how to systematically sample from a pool of disparate forecasts to create MFVs that effectively facilitate decision-making requires further investigation. To address this challenge, we examine two cluster-based sampling strategies for creating MFVs and three designs for visualizing them to assist people in decision-making with forecasts. Through two online studies (Experiment 1 n = 711 and Experiment 2 n = 400) and over 15 decision-making-related metrics, we evaluated participants’ perceptions of eight visualization designs using historical COVID-19 forecasts as a test bed. Our findings revealed that one sampling method significantly enhanced participants’ ability to predict future outcomes, thereby reducing their surprise when confronted with the actual outcomes. Importantly, since no approach excels in all metrics, we advise choosing different visualization designs based on communication goals. Furthermore, qualitative response data demonstrate a correlation between response consistency and people’s inclination to extrapolate from the forecast segment of the visualization. This research offers insights into how to improve visualizations of multiple forecasts using an automated and empirically validated technique for selecting forecasts that outperform common techniques on several key metrics and reduce overplotting.

Post-surgery care involves ongoing collaboration between provider teams and patients, which starts from post-surgery hospitalization through home recovery after discharge. While prior HCI research has primarily examined patients’ challenges at home, less is known about how provider teams coordinate discharge preparation and care handoffs, and how breakdowns in communication and care pathways may affect patient recovery. To investigate this gap, we conducted semi-structured interviews with 13 healthcare providers and 4 patients in the context of gastrointestinal (GI) surgery. We found coordination boundaries between in- and out-patient teams, coupled with complex organizational structures within teams, impeded the "invisible work" of preparing patients’ home care plans and triaging patient information. For patients, these breakdowns resulted in inadequate preparation for home transition and fragmented self-collected data, both of which undermine timely clinical decision-making. Based on these findings, we outline design opportunities to formalize task ownership and handoffs, contextualize co-temporal signals, and align care plans with home resources.

Simulating step-wise human behavior with Large Language Models (LLMs) has become an emerging research direction, enabling applications in various practical domains. While prior methods, including prompting, supervised fine-tuning (SFT), and reinforcement learning (RL), have shown promise in modeling step-wise behavior, they primarily learn a population-level policy without conditioning on a user’s persona, yielding generic rather than personalized simulations. In this work, we pose a critical question: how can LLM agents better simulate personalized user behavior? We introduce Customer-R1, an RL-based method for personalized, step-wise user behavior simulation in online shopping environments. Our policy is conditioned on an explicit persona, and we optimize next-step rationale and action generation via action correctness reward signals. Experiments on the OPeRA dataset emonstrate that Customer-R1 not only significantly outperforms prompting and SFT-based baselines in next-action prediction tasks, but also better matches users’ action distribution, indicating higher fidelity in personalized behavior simulation.

The dark patterns, deceptive interface designs manipulating user behaviors, have been extensively studied for their effects on human decision-making and autonomy. Yet, with the rising prominence of LLM-powered GUI agents that automate tasks from high-level intents, understanding how dark patterns affect agents is increasingly important. We present a two-phase empirical study examining how agents, human participants, and human-AI teams respond to 16 types of dark patterns across diverse scenarios. Phase 1 highlights that agents often fail to recognize dark patterns, and even when aware, prioritize task completion over protective action. Phase 2 revealed divergent failure modes: humans succumb due to cognitive shortcuts and habitual compliance, while agents falter from procedural blind spots. Human oversight improved avoidance but introduced costs such as attentional tunneling and cognitive load. Our findings show neither humans nor agents are uniformly resilient, and collaboration introduces new vulnerabilities, suggesting design needs for transparency, adjustable autonomy, and oversight.

While just-in-time interventions (JITIs) have effectively targeted common health behaviors, individuals often have unique needs to intervene in personal undesirable actions that can negatively affect physical, mental, and social well-being. We present WatchGuardian, a smartwatch-based JITI system that empowers users to define custom interventions for these personal actions with a small number of samples. For the model to detect new actions based on limited new data samples, we developed a few-shot learning pipeline that finetuned a pre-trained inertial measurement unit (IMU) model on public hand-gesture datasets. We then designed a data augmentation and synthesis process to train additional classification layers for customization. Our offline evaluation with 26 participants showed that with three, five, and ten examples, our approach achieved an average accuracy of 76.8%, 84.7%, and 87.7%, and an F1 score of 74.8%, 84.2%, and 87.2% We then conducted a four-hour intervention study to compare WatchGuardian against a rule-based intervention. Our results demonstrated that our system led to a significant reduction by 64.0 +- 22.6% in undesirable actions, substantially outperforming the baseline by 29.0%. Our findings underscore the effectiveness of a customizable, AI-driven JITI system for individuals in need of behavioral intervention in personal undesirable actions. We envision that our work can inspire broader applications of user-defined personalized intervention with advanced AI solutions.

Surgical site infection (SSI) is one of the most common and costly healthcare-associated infections and and surgical wound care remains a significant clinical challenge in preventing SSIs and improving patient outcomes. While recent studies have explored the use of deep learning for preliminary surgical wound screening, progress has been hindered by concerns over data privacy and the high costs associated with expert annotation. Currently, no publicly available dataset or benchmark encompasses various types of surgical wounds, resulting in the absence of an open-source Surgical-Wound screening tool. To address this gap: (1) we present SurgWound, the first open-source dataset featuring a diverse array of surgical wound types. It contains 697 surgical wound images annotated by 3 professional surgeons with eight fine-grained clinical attributes. (2) Based on SurgWound, we introduce the first benchmark for surgical wound diagnosis, which includes visual question answering (VQA) and report generation tasks to comprehensively evaluate model performance. (3) Furthermore, we propose a three-stage learning framework, WoundQwen, for surgical wound diagnosis. In the first stage, we employ five independent MLLMs to accurately predict specific surgical wound characteristics. In the second stage, these predictions serve as additional knowledge inputs to two MLLMs responsible for diagnosing outcomes, which assess infection risk and guide subsequent interventions. In the third stage, we train a MLLM that integrates the diagnostic results from the previous two stages to produce a comprehensive report. This three-stage framework can analyze detailed surgical wound characteristics and provide subsequent instructions to patients based on surgical images, paving the way for personalized wound care, timely intervention, and improved patient outcomes.

Metaphors play a critical role in shaping user experiences with Voice User Interfaces (VUIs), yet existing designs often rely on static, human-centric metaphors that fail to adapt to diverse contexts and user needs. This paper introduces Metaphor-Fluid Design, a novel approach that dynamically adjusts metaphorical representations based on conversational use-contexts. We compare this approach to a Default VUI, which characterizes the present implementation of commercial VUIs commonly designed around the persona of an assistant, offering a uniform interaction style across contexts. In Study 1 (N=130), metaphors were mapped to four key use-contexts-commands, information seeking, sociality, and error recovery-along the dimensions of formality and hierarchy, revealing distinct preferences for task-specific metaphorical designs. Study 2 (N=91) evaluates a Metaphor-Fluid VUI against a Default VUI, showing that the Metaphor-Fluid VUI enhances perceived intention to adopt, enjoyment, and likability by aligning better with user expectations for different contexts. However, individual differences in metaphor preferences highlight the need for personalization. These findings challenge the one-size-fits-all paradigm of VUI design and demonstrate the potential of Metaphor-Fluid Design to create more adaptive and engaging human-AI interactions.

The advancements of Large Language Models (LLMs) have decentralized the responsibility for the transparency of AI usage. Specifically, LLM users are now encouraged or required to disclose the use of LLM-generated content for varied types of real-world tasks. However, an emerging phenomenon, users’ secret use of LLM , raises challenges in ensuring end users adhere to the transparency requirement. Our study used mixed-methods with an exploratory survey (125 real-world secret use cases reported) and a controlled experiment among 300 users to investigate the contexts and causes behind the secret use of LLMs. We found that such secretive behavior is often triggered by certain tasks, transcending demographic and personality differences among users. Task types were found to affect users’ intentions to use secretive behavior, primarily through influencing perceived external judgment regarding LLM usage. Our results yield important insights for future work on designing interventions to encourage more transparent disclosure of the use of LLMs or other AI technologies.

Usability testing is a fundamental yet challenging research method for user experience (UX) researchers to evaluate a web design. Recent advances in Large Language Model-simulated Agent (LLM Agent) research inspired us to design UXAgent to support UX researchers in evaluating and reiterating their usability testing study design before they conduct the real human-subject study. Our system features an LLM Agent module and a universal browser connector module so that UX researchers can automatically generate thousands of simulated users to test the target website. The system can generate UX study results in qualitative (e.g., interviewing how an agent thinks), quantitative (e.g., # of actions), and video recording formats for UX researchers to analyze. Through a heuristic user evaluation with five UX researchers, participants praised the innovation of our system but also expressed concerns about the future of UX study with LLM Agents1.

In this paper, we study whether language models are affected by learned gender stereotypes during the comprehension of stories. Specifically, we investigate how models respond to gender stereotype perturbations through counterfactual data augmentation. Focusing on Question Answering (QA) tasks in fairytales, we modify the FairytaleQA dataset by swapping gendered character information and introducing counterfactual gender stereotypes during training. This allows us to assess model robustness and examine whether learned biases influence story comprehension. Our results show that models exhibit slight performance drops when faced with gender perturbations in the test set, indicating sensitivity to learned stereotypes. However, when fine-tuned on counterfactual training data, models become more robust to anti-stereotypical narratives. Additionally, we conduct a case study demonstrating how incorporating counterfactual anti-stereotype examples can improve inclusivity in downstream applications.

Older adult patients constitute a rapidly growing subgroup of Intensive Care Unit (ICU) patients. In these situations, their family caregivers are expected to represent the unconscious patients to access and interpret patients’ medical information. However, caregivers currently have to rely on overloaded clinicians for information updates and typically lack the health literacy to understand complex medical information. Our project aims to explore the information needs of caregivers of ICU older adult patients, from which we can propose design opportunities to guide future AI systems. The project begins with formative interviews with 11 caregivers to identify their challenges in accessing and interpreting medical information; From these findings, we then synthesize design requirements and propose an AI system prototype to cope with caregivers’ challenges. The system prototype has two key features: a timeline visualization to show the AI extracted and summarized older adult patients’ key medical events; and an LLM-based chatbot to provide context-aware informational support. We conclude our paper by reporting on the follow-up user evaluation of the system and discussing future AI-based systems for ICU caregivers of older adults.

Interactive story reading is a common parent-child activity, where parents expect to teach both language skills and real-world knowledge beyond the story. While increasing storytelling and reading systems have been developed for this activity, they often fail to infuse real-world knowledge into the conversation. This limitation can be attributed to the existing question-answering (QA) datasets used for children’s education, upon which the systems are built, failing to capture the nuances of how education experts think when conducting interactive story reading activities. To bridge this gap, we design an annotation framework, empowered by existing knowledge graph to capture experts’ annotations and thinking process, and leverage this framework to construct StorySparkQA dataset, which comprises 5, 868 expert-annotated QA pairs with real-world knowledge. We conduct automated and human expert evaluations across various QA pair generation settings to demonstrate that our StorySparkQA can effectively support models in generating QA pairs that target real-world knowledge beyond story content. StorySparkQA is available at https://huggingface.co/datasets/NEU-HAI/StorySparkQA.

Interactive storytelling is vital for preschooler development. While children’s interactive partners have traditionally been their parents and teachers, recent advances in artificial intelligence (AI) have sparked a surge of AI-based storytelling technologies. As these technologies become increasingly ubiquitous in preschoolers’ lives, questions arise regarding how they function in practical storytelling scenarios and, in particular, how parents, the most critical stakeholders, experience and perceive these technologies. This paper investigates these questions through a qualitative study with 17 parents of children aged 3-6. Our findings suggest that even though AI-based storytelling technologies provide more immersive and engaging interaction, they still cannot meet parents’ expectations due to a series of interactive, functional, and algorithmic challenges. We elaborate on these challenges and discuss the possible implications of future AI-based storytelling technologies for preschoolers. We conclude by highlighting the design implications for future AI-based storytelling technologies.

AI assistance in decision-making has become popular, yet people’s inappropriate reliance on AI often leads to unsatisfactory human-AI collaboration performance. In this paper, through three pre-registered, randomized human subject experiments, we explore whether and how the provision of second opinions may affect decision-makers’ behavior and performance in AI-assisted decision-making. We find that if both the AI model’s decision recommendation and a second opinion are always presented together, decision-makers reduce their over-reliance on AI while increase their under-reliance on AI, regardless whether the second opinion is generated by a peer or another AI model. However, if decision-makers have the control to decide when to solicit a peer’s second opinion, we find that their active solicitations of second opinions have the potential to mitigate over-reliance on AI without inducing increased under-reliance in some cases. We conclude by discussing the implications of our findings for promoting effective human-AI collaborations in decision-making.

Despite the promises of ML in education, its adoption in the classroom has surfaced numerous issues regarding fairness, accountability, and transparency, as well as concerns about data privacy and student consent. A root cause of these issues is the lack of understanding of the complex dynamics of education, including teacher-student interactions, collaborative learning, and classroom environments. To overcome these challenges and fully utilize the potential of ML in education, software practitioners need to work closely with educators and students to fully understand the context of the data (the backbone of ML applications) and collaboratively define the ML data specifications. To gain a deeper understanding of such a collaborative process, we conduct ten co-design sessions with ML software practitioners, educators, and students. In the sessions, teachers and students work with ML engineers, UX designers, and legal practitioners to define dataset characteristics for a given ML application. We find that stakeholders contextualize data based on their domain and procedural knowledge, proactively design data requirements to mitigate downstream harms and data reliability concerns, and exhibit role-based collaborative strategies and contribution patterns. Further, we find that beyond a seat at the table, meaningful stakeholder participation in ML requires structured supports: defined processes for continuous iteration and co-evaluation, shared contextual data quality standards, and information scaffolds for both technical and non-technical stakeholders to traverse expertise boundaries.

While most existing works on LLM prompt-engineering focus only on how to select a better set of data samples inside one single prompt input (In-Context Learning or ICL), why can’t we design and leverage multiple prompt inputs together to further improve the LLM performance? In this work, we propose In-Context Sampling (ICS), a low-resource LLM prompt-engineering technique to produce the most confident prediction results by optimizing the construction of multiple ICL prompt inputs. Extensive experiments with two SOTA LLMs (FlanT5-XL and Mistral-7B) on three NLI datasets (e-SNLI, Multi-NLI, and ANLI) illustrate that ICS can consistently enhance LLM’s prediction performance and confidence. An ablation study suggests that a diversity-based ICS strategy may further improve LLM’s performance, which sheds light on a new yet promising future research direction.

Research into AI alignment has grown considerably since the recent introduction of increasingly capable Large Language Models (LLMs). Unfortunately, modern methods of alignment still fail to fully prevent harmful responses when models are deliberately attacked. Such vulnerabilities can lead to LLMs being manipulated into generating hazardous content: from instructions for creating dangerous materials to inciting violence or endorsing unethical behaviors. To help mitigate this issue, we introduce Bergeron: a framework designed to improve the robustness of LLMs against attacks without any additional parameter fine-tuning. Bergeron is organized into two tiers; with a secondary LLM acting as a guardian to the primary LLM. This framework better safeguards the primary model against incoming attacks while monitoring its output for any harmful content. Empirical analysis reviews that by using Bergeron to complement models with existing alignment training, we can significantly improve the robustness and safety of multiple, commonly used commercial and open-source LLMs. Specifically, we found that models integrated with Bergeron are, on average, nearly seven times more resistant to attacks compared to models without such support.

Cardiotoxicity induced by cancer treatment has become a major clinical concern, affecting the long-term survival and quality of life of cancer patients. Effective clinical decision-making, including the detection of cancer treatment-induced cardiotoxicity and the monitoring of associated symptoms, remains a challenging task for clinicians. This study investigates the current practices and needs of clinicians in the clinical decision making of cancer treatment-induced cardiotoxicity and explores the potential of digital health technologies to support this process. Through semi-structured interviews with seven clinical experts, we identify a three-step decision-making paradigm: 1) symptom identification, 2) diagnostic testing and specialist collaboration, and 3) clinical decision-making and intervention. Our findings highlight the difficulties of diagnosing cardiotoxicity (absence of unified protocols and high variability in symptoms) and monitoring patient symptoms (lacking accurate and timely patient self-reported symptoms). The clinicians also expressed their need for effective early detection tools that can integrate remote patient monitoring capabilities. Based on these insights, we discuss the importance of understanding the dynamic nature of clinical workflows, and the design considerations for future digital tools to support cancer-treatment-induced cardiotoxicity decision-making.

Theory of Mind (ToM) significantly impacts human collaboration and communication as a crucial capability to understand others. When AI agents with ToM capability collaborate with humans, Mutual Theory of Mind (MToM) arises in such human-AI teams (HATs). The MToM process, which involves interactive communication and ToM-based strategy adjustment, affects the team’s performance and collaboration process. To explore the MToM process, we conducted a mixed-design experiment using a large language model-driven AI agent with ToM and communication modules in a real-time shared-workspace task. We find that the agent’s ToM capability does not significantly impact team performance but enhances human understanding of the agent and the feeling of being understood. Most participants in our study believe verbal communication increases human burden, and the results show that bidirectional communication leads to lower HAT performance. We discuss the results’ implications for designing AI agents that collaborate with humans in real-time shared workspace tasks.