Lingrui Mei, Shenghua Liu, Yiwei Wang, Yuyao Ge, Baolong Bi, Jiayu Yao, Jun Wan, Ziling Yin, Jiafeng Guo, Xueqi Cheng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Oral health directly influences children's overall well-being, yet pervasive dental anxiety has become an unavoidable barrier to pediatric dental care.While behavior management proves more effective than environmental or equipment improvements in reducing pediatric dental anxiety while improving treatment understanding and oral health awareness, its time-intensive nature often conflicts with dentists' demanding workloads.Through formative research, we proposed a structured combination of behavior management techniques within dental waiting rooms, employing tangible interactions to create a comprehensive anxiety relief system for children aged 5-10 years.The system encompasses two core processes, dental caries treatment and caries prevention.We conducted the validation experiment and pilot study to refine the system.Then we deployed a user study involving 32 child-parent groups.The results demonstrated that the system effectively alleviates children's dental anxiety, prepares them for dental visits, promotes parent-child interaction, and supports children's participation.

Multimodal Retrieval-Augmented Generation (RAG) systems have become essential in knowledge-intensive and open-domain tasks.As retrieval complexity increases, ensuring the robustness of these systems is critical.However, current RAG models are highly sensitive to the order in which evidence is presented, often resulting in unstable performance and biased reasoning, particularly as the number of retrieved items or modality diversity grows.This raises a central question: How does the position of retrieved evidence affect multimodal RAG performance?To answer this, we present the first comprehensive study of position bias in multimodal RAG systems.Through controlled experiments across text-only, imageonly, and mixed-modality tasks, we observe a consistent U-shaped accuracy curve with respect to evidence position.To quantify this bias, we introduce the Position Sensitivity Index (P SI p ) and develop a visualization framework to trace attention allocation patterns across decoder layers.Our results reveal that multimodal interactions intensify position bias compared to unimodal settings, and that this bias increases logarithmically with retrieval range.These findings offer both theoretical and empirical foundations for position-aware analysis in RAG, highlighting the need for evidence reordering or debiasing strategies to build more reliable and equitable generation systems.Our code and experimental resources are available at https://github.com/Theodyy/ Multimodal-Rag-Position-Bias.

The surge in AI-generated images, blending authentic and synthetic content, raises security concerns and complicates model attribution, especially with limited transparency. Existing methods either struggle to attribute across multiple frameworks or rely on additional conditions, such as textual descriptions and white-box access to the source model, limiting their practicality and effectiveness in real-world scenarios. To address this gap, we introduce Model Attribution via Inverse Diffusion (MAID), the first framework-agnostic and self-sufficient approach that leverages the source model features extracted by diffusion models, which also works for images generated from GANs. By employing the inverse diffusion process, we are able to utilize pre-trained Diffusion Models as Denoising Autoencoders, mapping images into a latent space and extracting the Diffusion Model Activations (DMA). This mapping effectively captures the unique characteristics of images originating from different source models, including authentic images, which showcase distinct latent Gaussian signatures. Experimental results show that, even in data-asymmetric unfair comparisons, the attribution classifier trained with our proposed DMA achieves approximately 15% and 3% higher ACC compared to SOTA methods on the DiffusionForensics and Artifact datasets, respectively. The code is available at https://github.com/Zhu-Luyu/MAID.

Picky eating behavior is common in preschoolers and has been linked to a lack of food literacy with support from certain research. Recent research has focused on interventions for children’s mealtime behaviors which can lead to distraction and neglect of food literacy learning. We propose FeastyMaze, a tangible and multi-sensory interactive puzzle game for young children to improve eating behavior. With the Five-color Diet Theory, our approach enables children to actively learn about food nutrition and balanced diets. To evaluate the effectiveness and acceptability of FeastyMaze, we conducted a user study with preschoolers (N = 12) who exhibited picky eating behaviors. The results showed that it effectively increased children’s familiarity and understanding of food knowledge, built positive attitudes towards previously disliked foods, and had the potential to improve their eating behavior.

Laser cutting revolutionizes the creation of personal-fabricated prototypes. These objects can have transformable properties by adopting different materials and be interactive by integrating electronic circuits. However, circuits in laser-cut objects always have limited movements, which refrains laser cutting from achieving interactive prototypes with more complex movable functions like mechanisms. We propose MechCircuit, a design and fabrication pipeline for making mechanical-electronical objects with laser cutting. We leverage the neodymium magnet’s natures of magnetism and conductivity to integrate electronics and mechanical structure joints into prototypes. We conduct the evaluation to explore technological parameters and assess the practical feasibility of the fabrication pipeline. And we organized a user-observing workshop for non-expert users. Through the outcoming prototypes, the result demonstrates the feasibility of MechCircuit as a useful and inspiring prototyping method.

The amount and quality of mathematical language in the family are positively associated with promoting children’s mathematical abilities. However, mathematical language in many families is poor. Through need-finding investigation, we developed MathKingdom, a voice-agent-based game that helps children aged 4–7 learn and use rich, accurate mathematical language (e.g., mathematical expressions related to measurement, sequence, patterns). The game has four flows, in which users can wake up, transform, decorate, and perform as their avatars, as well as practice basic mathematical vocabulary, mathematical single sentences, coherent mathematical statements, and free expression. We refined the system design through wizard-of-oz testing and then evaluated it with 18 families. The results showed that MathKingdom effectively engaged children, enhanced their mathematical language skills and mathematical abilities, and encouraged parent-child conversations about math.

Most intervention methods used for social skills training in children with autism are dedicated to high-functioning autism (HFA). However, extensive neurological and developmental disorders of low-functioning autism (LFA) have hampered their adoption. In this study, we observed and interviewed children with LFA, and their teachers, from a local educational institution, to better understand the children's social needs and barriers. Then, with the aim of aiding the children with their social activities, we illustrate the design process of SqueeBall, an interactive playground equipment. We evaluated the design with 18 children (16 with LFA and 2 with HFA) between 2.5 and 7 years of age. Results showed that these children had a pleasant game experience when the group bonded, and the equipment had a positive effect on aiding them in various ways. Finally, we discuss the challenges and opportunities of multimedia interaction techniques in aiding children with LFA.

The promise of augmenting accurate predictions provided by modern neural networks with well-calibrated predictive uncertainties has reinvigorated interest in Bayesian neural networks. However, model selection---even choosing the number of nodes---remains an open question. Poor choices can severely affect the quality of the produced uncertainties. In this paper, we explore continuous shrinkage priors, the horseshoe, and the regularized horseshoe distributions, for model selection in Bayesian neural networks. When placed over node pre-activations and coupled with appropriate variational approximations, we find that the strong shrinkage provided by the horseshoe is effective at turning off nodes that do not help explain the data. We demonstrate that our approach finds compact network structures even when the number of nodes required is grossly over-estimated. Moreover, the model selection over the number of nodes does not come at the expense of predictive or computational performance; in fact, we learn smaller networks with comparable predictive performance to current approaches. These effects are particularly apparent in sample-limited settings, such as small data sets and reinforcement learning.

Bayesian Neural Networks (BNNs) have recently received increasing attention for their ability to provide well-calibrated posterior uncertainties. However, model selection—even choosing the number of nodes—remains an open question. Recent work has proposed the use of a horseshoe prior over node pre-activations of a Bayesian neural network, which effectively turns off nodes that do not help explain the data. In this work, we propose several modeling and inference advances that consistently improve the compactness of the model learned while maintaining predictive performance, especially in smaller-sample settings including reinforcement learning.

Aggregate information is important for data stream processing systems, especially, to get any user-specified time window queries and evolution analysis over data stream. To solve this problem, in this paper, we propose an integrated structure for managing summarized information of snapshots under geometry timeframe, which facilitates analyzing the temporal evolving of data stream. By using our method, the cost of update operations and the errors can be controlled within acceptable scope. Evaluation shows the our structure can respond to arbitrary time window aggregate queries within small errors, efficiently.