Large language model (LLM) agents are increasingly acting as human delegates in multiagent environments, where a representative agent integrates diverse peer perspectives to make a final decision.Drawing inspiration from social psychology, we investigate how the reliability of this representative agent is undermined by the social context of its network.We define four key phenomena-social conformity, perceived expertise, dominant speaker effect, and rhetorical persuasion-and systematically manipulate the number of adversaries, relative intelligence, argument length, and argumentative styles.Our experiments demonstrate that the representative agent's accuracy consistently declines as social pressure increases: larger adversarial groups, more capable peers, and longer arguments all lead to significant performance degradation.Furthermore, rhetorical strategies emphasizing credibility or logic can further sway the agent's judgment, depending on the context.These findings reveal that multiagent systems are sensitive not only to individual reasoning but also to the social dynamics of their configuration, highlighting critical vulnerabilities in AI delegates that mirror the psychological biases observed in human group decision-making.

Junmyeong Lee, Chan Hur, ChangSu Choi, Sukmin Cho, Fitsum Gaim, Eui Jun Hwang, Hoyun Song, KyungTae Lim. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Eui Jun Hwang, Sukmin Cho, Junmyeong Lee, Jong C. Park. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Content moderation research has recently made significant advances, but remains limited in serving the majority of the world's languages due to the lack of resources, leaving millions of vulnerable users to online hostility. This work presents a large-scale human-annotated multi-task benchmark dataset for abusive language detection in Tigrinya social media with joint annotations for three tasks: abusiveness, sentiment, and topic classification. The dataset comprises 13,717 YouTube comments annotated by nine native speakers, collected from 7,373 videos with a total of over 1.2 billion views across 51 channels. We developed an iterative term clustering approach for effective data selection. Recognizing that around 64% of Tigrinya social media content uses Romanized transliterations rather than native Ge'ez script, our dataset accommodates both writing systems to reflect actual language use. We establish strong baselines across the tasks in the benchmark, while leaving significant challenges for future contributions. Our experiments demonstrate that small fine-tuned models outperform prompted frontier large language models (LLMs) in the low-resource setting, achieving 86.67% F1 in abusiveness detection (7+ points over best LLM), and maintain stronger performance in all other tasks. The benchmark is made public to promote research on online safety.

This work addresses the challenges associated with the use of glosses in both Sign Language Translation (SLT) and Sign Language Production (SLP). While glosses have long been used as a bridge between sign language and spoken language, they come with two major limitations that impede the advancement of sign language systems. First, annotating the glosses is a labor-intensive and time-consuming process, which limits the scalability of datasets. Second, the glosses oversimplify sign language by stripping away its spatio-temporal dynamics, reducing complex signs to basic labels and missing the subtle movements essential for precise interpretation. To address these limitations, we introduce Universal Gloss-level Representation (UniGloR), a framework designed to capture the spatio-temporal features inherent in sign language, providing a more dynamic and detailed alternative to the use of the glosses. The core idea of UniGloR is simple yet effective: We derive dense spatiotemporal representations from sign keypoint sequences using self-supervised learning and seamlessly integrate them into SLT and SLP tasks. Our experiments in a keypoint-based setting demonstrate that UniGloR either outperforms or matches the performance of previous SLT and SLP methods on two widely-used datasets: PHOENIX14T and How2Sign. Code is available at https://github.com/eddie-euijun-hwang/UniGloR.

Gloss-free Sign Language Production (SLP) offers a direct translation of spoken language sentences into sign language, bypassing the need for gloss intermediaries. Previous autoregressive SLP methods have not fully achieved true autoregression, as they often depend on ground-truth data during inference. To fill this gap, we introduce Sign language Vector Quantization Network (SignVQNet), leveraging discrete spatio-temporal representations of sign poses. With such a discrete representation, our method incorporates beam search, a decoding strategy widely used in Natural Language Processing. Furthermore, we align the discrete representation with linguistic features from pre-trained language models such as BERT. Our results show the superior performance of our method over prior SLP methods in generating accurate and realistic sign pose sequences. Additionally, our analysis shows that the reliability of Back-Translation and Fréchet Gesture Distance as evaluation metrics, in contrast to DTW-MJE. The code and models are available at https://github.com/eddie-euijun-hwang/SignVQNet.

Sign language production (SLP) is the process of generating sign language videos from spoken language expressions. Since sign languages are highly under-resourced, existing vision-based SLP approaches suffer from out-of-vocabulary (OOV) and test-time generalization problems and thus generate low-quality translations. To address these problems, we introduce an avatar-based SLP system composed of a sign language translation (SLT) model and an avatar animation generation module. Our Transformer-based SLT model utilizes two additional strategies to resolve these problems: named entity transformation to reduce OOV tokens and context vector generation using a pretrained language model (e.g., BERT) to reliably train the decoder. Our system is validated on a new Korean-Korean Sign Language (KSL) dataset of weather forecasts and emergency announcements. Our SLT model achieves an 8.77 higher BLEU-4 score and a 4.57 higher ROUGE-L score over those of our baseline model. In a user evaluation, 93.48% of named entities were successfully identified by participants, demonstrating marked improvement on OOV issues.

Task-oriented dialogue system has a vital role in Human-Robot Interaction (HRI). However, it has been developed based on conventional pipeline approach which has several drawbacks; expensive, time-consuming, and so on. Based on this approach, developers manually define a robot's behaviour such as gestures and facial expressions on the corresponding dialogue states. Recently, end-to-end learning of Recurrent Neural Networks (RNNs) is an attractive solution for the dialogue system. In this paper, we proposed a social robot system using end-to-end dialogue system in the context of hospital receptionist. We utilized Hybrid Code Network (HCN) as an end-to-end dialogue system and extended to select both response and gesture using RNN based gesture selector. We evaluate its performance with human users and compare the results with one of the conventional methods. Empirical result shows that the proposed method has benefits in terms of dialogue efficiency, which indicates how efficient users were in performing the given tasks with the help of the robot. Moreover, we achieved the same performance regarding the robot's gesture with the proposed method compared to manually defined gestures.

This paper presents a new version of robot receptionist system for healthcare facility environment. Our HealthBots consists of three subsystems: a receptionist robot system, a nurse assistant robot system, and a medical server. Our first version of receptionist robot, interacts with human at hospital reception, gives instructions to human verbally, but cannot understand what human says, so it uses a touch screen to get the response from human. In this paper, we design a receptionist robot that recognizes human face as well as speech, which enhances verbal interaction skill of robot. In addition, we design a reaction generation engine to generate appropriate reactive motions and speech. Moreover, we study which social skills are important to a hospital receptionist robot to enhance social interaction, such as friendliness and attention. We implemented perception modules, decision-making modules, and reaction modules to our HealthBots architecture, and did two case studies to find essential social skills for hospital receptionist robots.