publications

2026

1. Deep learning characterizes depression and suicidal ideation in young adults from eye movements

   Kleanthis Avramidis, Woojae Jeong, Aditya Kommineni, and 14 more authors

   npj Digital Medicine, Mar 2026

   Abs arXiv

   Objective biobehavioral markers for mental health conditions remain elusive, with diagnosis typically relying on self-reports and clinical interviews. We investigate eye tracking as a potential marker of attentional and mood biases associated with symptoms of depression and suicidal ideation from self-reported screening questionnaires. We analyze eye movements from 126 young adults during reading and responding to emotionally loaded sentences. A deep learning framework was designed to account for intra-trial and inter-trial variations in eye movements, achieving an AUC of 0.793 (95% CI: 0.766–0.819) for identifying depression/suicidality against healthy controls, and 0.826 (95% CI: 0.798–0.853) for suicidality specifically. The model also exhibited moderate accuracy in differentiating depressed from suicidal individuals (AUC: 0.609, 95% CI: 0.569–0.646). Discriminative patterns were more pronounced during response generation and for stimuli of negative sentiment. These findings suggest that eye tracking can provide objective markers of self-reported symptom severity by measuring the impact of emotional stimuli on oculomotor control.

2. Intelligence Requires Grounding But Not Embodiment

   Marcus Ma, and Shrikanth Narayanan

   Mar 2026

   Abs arXiv

   Recent advances in LLMs have reignited scientific debate over whether embodiment is necessary for intelligence. We present the argument that intelligence requires grounding, a phenomenon entailed by embodiment, but not embodiment itself. We define intelligence as the possession of four properties—motivation, predictive ability, understanding of causality, and learning from experience—and argue that each can be achieved by a non-embodied, grounded agent. We use this to conclude that grounding, not embodiment, is necessary for intelligence. We then present a thought experiment of an intelligent LLM agent in a digital environment and address potential counterarguments.

3. Authors Should Label Their Own Documents

   Marcus Ma, Cole Johnson, Nolan Bridges, and 3 more authors

   Mar 2026

   Abs arXiv

   Third-party annotation is the status quo for labeling text, but egocentric information such as sentiment and belief can at best only be approximated by a third-person proxy. We introduce author labeling, an annotation technique where the writer of the document itself annotates the data at the moment of creation. We collaborate with a commercial chatbot with over 20,000 users to deploy an author labeling annotation system. We train and deploy an online-learning model architecture for product recommendation with author-labeled data and achieve a 537% improvement in click-through rate compared to an industry advertising baseline running concurrently. We then compare the quality and practicality of author labeling to three traditional annotation approaches for sentiment analysis and find author labeling to be higher quality, faster to acquire, and cheaper.

4. Encoding Emotion Through Self-Supervised Eye Movement Reconstruction

   Marcus Ma, Jordan Prescott, Emily Zhou, and 4 more authors

   In ICASSP 2026, Mar 2026

   Abs arXiv

   The relationship between emotional expression and eye movement is well-documented, with literature establishing gaze patterns are reliable indicators of emotion. However, most studies utilize specialized, high-resolution eye-tracking equipment, limiting the potential reach of findings. We investigate how eye movement can be used to predict multimodal markers of emotional expression from naturalistic, low-resolution videos. We utilize a collection of video interviews from the USC Shoah Foundation’s Visual History Archive with Holocaust survivors as they recount their experiences in the Auschwitz concentration camp. Inspired by pretraining methods on language models, we develop a novel gaze detection model that uses self-supervised eye movement reconstruction that can effectively leverage unlabeled video. We use this model’s encoder embeddings to fine-tune models on two downstream tasks related to emotional expression and conclude that self-supervised eye movement reconstruction is an effective method for encoding the affective signal they carry.

5. Looking Into the Past: Eye Movements Characterize Elements of Autobiographical Recall in Interviews with Holocaust Survivors

   Emily Zhou, Marcus Ma, Kleanthis Avramidis, and 2 more authors

   Mar 2026

   Abs arXiv

   Eye movement and memory retrieval are deeply and bidirectionally intertwined, however existing literature is generally confined to controlled lab settings. We investigate the relationship between eye gaze and memory recall in free-form autobiographical recall using a large video corpus of semi-naturalistic interviews with Holocaust survivors (N=806). We examine eye movements with respect to episodic, semantic, affective, and temporal dimensions of traumatic and highly emotional autobiographical recall. We observe gaze patterns vary significantly across certain temporal contexts, most prominently in vertical eye movements, and find that eye movements entirely preceding sentence onset are sufficient for predicting temporal context.

6. Smiling Regulates Emotion During Traumatic Recollection

   Marcus Ma, Emily Zhou, Leonard Ludwig, and 7 more authors

   In Proceedings of the 28th ACM International Conference on Multimodal Interaction, Mar 2026

   Oral presentation. Best Paper Runner-Up.

   Abs arXiv

   We study when, where, and why 978 Holocaust survivors smile in video testimonies. We create an automatic smile detection model from facial features with an F1 of 85% and annotate detected smiles under two established taxonomies of smiling. We produce narrative features on 1,083,417 transcript sentences as well as emotional valence from three different modalities: audio, eye gaze, and text transcript. Smiling rates are associated with specific semantic topics, narrative structures, and temporal syntaxes across the corpus. Smiles often occur during periods of intense negative affect and we find negative-affect smiles are associated with more positive subsequent valence trajectories across all three modalities. Smiling reduces eye dynamics and blink rates, with both of these effects modulated by narrative valence. Taken together, we conclude that smiling plays a critical role in regulating emotion and social interaction during traumatic recollection.

2025

1. CROSSNEWS: A Cross-Genre Authorship Verification and Attribution Benchmark

   Marcus Ma, Duong Minh Le, Junmo Kang, and 5 more authors

   Proceedings of the AAAI Conference on Artificial Intelligence, Apr 2025
2. “Crying like a Baby”: Survival Analysis and the Multimodal Memory of Holocaust Survivors

   Gabor Mihaly Toth, Mohamed Laib, Alina Bothe, and 4 more authors

   Anthology of Computers and the Humanities, Apr 2025
3. Understanding Human-AI Misalignment in LLM-Based Job-Seeking Support for Neurodivergent Users

   Kaely Hall, Marcus Ma, Xinyue Zhang, and 2 more authors

   In Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility, Apr 2025

   Abs

   Large Language Models are often trained on data reflecting neurotypical norms, yet are increasingly deployed to support neurodivergent users in sensitive domains like job-seeking. We examine interactions between neurodivergent job-seekers and a GPT-4-powered career support chatbot through the lens of misalignment. Through analysis of over 300 chat logs and interviews with 15 neurodivergent participants, we found that the chatbot frequently misrepresented users’ skills, imposed neurotypical language and expectations, and provided generic or inappropriate advice– even when relevant user data was available. Participants expected the chatbot to interpret implicit insights from their data, however, they sometimes lacked the clarity or confidence to correct the system when it did not, revealing gaps in both AI design and user understanding of system function. Our findings underscore the need for bi-directional alignment between neurodivergent users and LLMs, and call for design strategies that integrate neurodivergent perspectives and preferences to ensure more authentic, personalized, and human-centered AI support.

4. Humans Hallucinate Too: Language Models Identify and Correct Subjective Annotation Errors With Label-in-a-Haystack Prompts

   Georgios Chochlakis, Peter Wu, Tikka Arjun Singh Bedi, and 3 more authors

   In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Nov 2025

   Abs

   Modeling complex subjective tasks in Natural Language Processing, such as recognizing emotion and morality, is considerably challenging due to significant variation in human annotations. This variation often reflects reasonable differences in semantic interpretations rather than mere noise, necessitating methods to distinguish between legitimate subjectivity and error.We address this challenge by exploring label verification in these contexts using Large Language Models (LLMs). First, we propose a simple In-Context Learning binary filtering baseline that estimates the reasonableness of a document-label pair. We then introduce the Label-in-a-Haystack setting: the query and its label(s) are included in the demonstrations shown to LLMs, which are prompted to predict the label(s) again, while receiving task-specific instructions (e.g., emotion recognition) rather than label copying.We show how the failure to copy the label(s) to the output of the LLM are task-relevant and informative. Building on this, we propose the Label-in-a-Haystack Rectification (LiaHR) framework for subjective label correction: when the model outputs diverge from the reference gold labels, we assign the generated labels to the example instead of discarding it. This approach can be integrated into annotation pipelines to enhance signal-to-noise ratios. Comprehensive analyses, human evaluations, and ecological validity studies verify the utility of LiaHR for label correction. Code is available at https://github.com/gchochla/liahr.

5. Large Language Models Do Multi-Label Classification Differently

   Marcus Ma, Georgios Chochlakis, Niyantha Maruthu Pandiyan, and 2 more authors

   In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Nov 2025

   Abs

   Multi-label classification is prevalent in real-world settings, but the behavior of Large Language Models (LLMs) in this setting is understudied. We investigate how autoregressive LLMs perform multi-label classification, focusing on subjective tasks, by analyzing the output distributions of the models at each label generation step. We find that the initial probability distribution for the first label often does not reflect the eventual final output, even in terms of relative order and find LLMs tend to suppress all but one label at each generation step. We further observe that as model scale increases, their token distributions exhibit lower entropy and higher single-label confidence, but the internal relative ranking of the labels improves. Finetuning methods such as supervised finetuning and reinforcement learning amplify this phenomenon. We introduce the task of distribution alignment for multi-label settings: aligning LLM-derived label distributions with empirical distributions estimated from annotator responses in subjective tasks. We propose both zero-shot and supervised methods which improve both alignment and predictive performance over existing approaches. We find one method – taking the max probability over all label generation distributions instead of just using the initial probability distribution – improves both distribution alignment and overall F1 classification without adding any additional computation.

6. 75-Speaker Annot-16: A benchmark dataset for speech articulatory rt-MRI annotation with articulator contours and phonetic alignment

   Xuan Shi, Yubin Zhang, Yijing Lu, and 6 more authors

   In Interspeech 2025, Nov 2025

2024

1. Time-varying functional connectivity predicts fluctuations in sustained attention in a serial tapping task

   Dolly T Seeburger, Nan Xu, Marcus Ma, and 4 more authors

   Cogn Affect Behav Neurosci, Jan 2024

   Abs

   The mechanisms for how large-scale brain networks contribute to sustained attention are unknown. Attention fluctuates from moment to moment, and this continuous change is consistent with dynamic changes in functional connectivity between brain networks involved in the internal and external allocation of attention. In this study, we investigated how brain network activity varied across different levels of attentional focus (i.e., “zones”). Participants performed a finger-tapping task, and guided by previous research, in-the-zone performance or state was identified by low reaction time variability and out-of-the-zone as the inverse. In-the-zone sessions tended to occur earlier in the session than out-of-the-zone blocks. This is unsurprising given the way attention fluctuates over time. Employing a novel method of time-varying functional connectivity, called the quasi-periodic pattern analysis (i.e., reliable, network-level low-frequency fluctuations), we found that the activity between the default mode network (DMN) and task positive network (TPN) is significantly more anti-correlated during in-the-zone states versus out-of-the-zone states. Furthermore, it is the frontoparietal control network (FPCN) switch that differentiates the two zone states. Activity in the dorsal attention network (DAN) and DMN were desynchronized across both zone states. During out-of-the-zone periods, FPCN synchronized with DMN, while during in-the-zone periods, FPCN switched to synchronized with DAN. In contrast, the ventral attention network (VAN) synchronized more closely with DMN during in-the-zone periods compared with out-of-the-zone periods. These findings demonstrate that time-varying functional connectivity of low frequency fluctuations across different brain networks varies with fluctuations in sustained attention or other processes that change over time.

2. Introducing Bed Word: a new automated speech recognition tool for sociolinguistic interview transcription

   Marcus Ma, Lelia Glass, and James Stanford

   Linguistics Vanguard, Jan 2024

2023

1. It Takes Two to Avoid Pregnancy: Addressing Conflicting Perceptions of Birth Control Pill Responsibility in Romantic Relationships

   Marcus Ma, Chae Hyun Kim, Kaely Hall, and 1 more author

   Proc. ACM Hum.-Comput. Interact., Oct 2023

   Abs

   While birth control pills are one of the most common forms of contraception, their usage has several emotional and physical costs, such as taking the pill daily and experiencing hormonal side effects. The burden of these tasks in relationships generally falls on the pill user with minimal involvement from their partner. In this study, we conducted semi-structured interviews with pill users and their partners to investigate the differences between their perceived current and ideal divisions of birth control responsibility. During the interview, we presented a collaborative birth control tracking app prototype to examine how such technology can overcome these discrepancies. We found that pill users were unsatisfied with their partners’ engagement in contraceptive tasks but did not communicate this well. Meanwhile, partners wanted to contribute more to pregnancy prevention but did not know how. When presented with our app prototype, users and partners stated that our design could address these issues by improving communication between users and partners. In particular, users appreciated how technology could increase engagement and support from their partner, and partners liked that our app presented several concrete ways to become more involved and show emotional support. However, privacy issues exist given the sensitive nature of contraception. We highlight design considerations that should be kept in mind about privacy while recognizing pill users’ efforts and promoting partners’ involvement.