User simulation has long been important in computer science because it helps researchers build and test many kinds of systems. Since language is the main way people communicate, being able to simulate conversations has become especially valuable. Recent advances in large language models, or LLMs, have made this much more realistic by allowing computers to generate synthetic user conversations that closely resemble human dialogue. This paper reviews recent progress in LLM-based conversational user simulation. It introduces a new framework for organizing the field based on how detailed the simulated user is and what the simulation is meant to achieve. The paper also examines the main methods used and how researchers evaluate these systems. Overall, it aims to give readers a clear view of the latest developments in conversational user simulation and to identify the main challenges that still need to be solved.

Fu, Haowei; Ni, Bo; Xu, Han; Liu, Kunpeng; Lin, Dan; Derr, Tyler. (2026).��.��In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics: Findings of EACL 2026, 2786–2799.��

This paper describes a method that uses large language models, or LLMs, to combine information from several sources and infer students’ metacognitive behaviors, meaning the ways they plan, monitor, reflect on, and adjust their own learning. The study analyzes multimodal classroom data, including students’ movements, gaze, gestures, and speech, collected during a mixed-reality simulation shown on a classroom screen. Instead of processing each type of data separately, the researchers use the LLM at a late-fusion stage, meaning after the different signals have already been analyzed, to bring them together through prompting strategies such as zero-shot prompting, self-consistency reasoning, and carefully designed prompts. They also test whether an “LLM-as-a-Judge,” an LLM used to evaluate model outputs, can reliably assess these behavior labels at scale and reduce the need for manual review. Using a balanced set of human-checked examples and control cases, they compare text-based LLMs, such as GPT-5, with visual-language models, or VLMs, such as Qwen2.5-VL, which can directly process images or video. The results show that text-based LLMs used in this late-fusion way can outperform VLMs even without raw video, and that prompt design can shift the model toward being more precise or more sensitive when the behavior is subtle or brief. Overall, the findings suggest that LLMs can be effective tools for combining multimodal learning data and that LLM-as-a-Judge can support scalable, human-in-the-loop evaluation.

Figure 1:

On-screen, each tracked student is represented as a molecule. As they move through the classroom space, their corresponding molecule moves in real time, allowing them to explore the photosynthesis cycle through embodied interaction.

]]> /valiant/2026/06/17/using-large-language-models-to-detect-socially-shared-regulation-of-collaborative-learning/ Wed, 17 Jun 2026 15:43:51 +0000 /valiant/?p=6966 Zhang, Jiayi; Borchers, Conrad; Cohn, Clayton; Srivastava, Namarata; Snyder, Caitlin; Guo, Siyuan; Ashwin, T. S.; Mohammed, Naveeduddin; Noh, Haley; Biswas, Gautam. (2026).��.��16th International Learning Analytics and Knowledge Conference, LAK 2026, 883–890.��

The field of learning analytics has made progress in automatically identifying complicated learning behaviors from different kinds of data, including text, audio, and system logs. But most of this work has focused on individual problem-solving, not on collaborative, open-ended problem-solving, where students work together on tasks and their behavior can be harder to interpret because group interactions are more varied and less predictable. This study extended predictive models to detect socially shared regulation of learning, or SSRL, which refers to the ways group members plan, monitor, and adjust their learning together, in collaborative computational modeling environments. The researchers used large language models, or LLMs, as summarization tools to turn student dialogue into task-relevant representations that could be matched with system log data. They compared several kinds of inputs, including text-only embeddings, which are numerical representations of the dialogue; context-rich embeddings, which include more information about the situation; and features derived from the logs. The results showed that text-only embeddings were often best at detecting SSRL behaviors tied to carrying out the task or group activity, such as going off task or asking for help. Meanwhile, the context-rich and multimodal features were especially useful for identifying planning and reflection. Overall, the study suggests that embedding-based models could make it possible to detect group learning behaviors at scale and eventually support real-time feedback and adaptive help in collaborative learning settings.

Figure 1:

Example solution for the C2STEM Truck Task with Task Context categories [].

This study looked at how students use AI chat agents in problem-based learning, or PBL, where learners solve open-ended problems by exploring information and ideas. The researchers focused on conversational agents, or CAs, powered by large language models (LLMs), which are AI systems that can generate human-like text. These agents were built to support different parts of inquiry in a story-based learning environment: one provided content knowledge, one gave feedback on arguments, and one evaluated arguments. Using log data from 15 student groups and Pedaste et al.’s inquiry cycle, a framework that describes the main stages of inquiry learning, the researchers examined how students interacted with each agent over time. They found that the agents played different but complementary roles: they could help students search for information, revise their ideas, and reflect on their thinking. At the same time, the agents sometimes steered students in ways that limited exploration. Overall, the study shows that learning analytics, which uses digital data to study learning behavior, can help educators understand how students work with AI support and design more flexible tools for classroom use.

Figure 1:

Example of game characters and conversational agent interface.

Skyler Grandel; Scott Thomas Andersen; Yu Huang; Kevin Leach (2026).��.��ACM Transactions on Software Engineering and Methodology, 35(3), Article 82.��

Large Language Models (LLMs) have been very successful at analyzing graphs—such as predicting node classification (labeling items in a network) and link prediction (predicting missing connections). However, little research has explored whether LLMs can actually generate new graph structures. This study investigates that question.

We designed prompts that guide LLMs to write code that creates graphs with specific structural properties, using ideas from network science. Different types of networks—such as social networks or transportation networks—have different structural patterns. For example, the clustering coefficient measures how often triangles appear in social networks, while square patterns may reflect road layouts in transportation systems. We first tested whether LLMs could generate graphs that match these kinds of domain-specific structural properties.

Next, we selected the best-performing configurations and compared LLM-generated graphs with those produced by established graph generative models across multiple domains. Our results provide insight into how well LLMs can generate realistic network structures and where their strengths and limitations lie.

The exponential growth of Large Language Models (LLMs) continues to highlight the need for efficient strategies to meet ever-expanding computational and data demands. This survey provides a comprehensive analysis of two complementary paradigms: Knowledge Distillation (KD) and Dataset Distillation (DD), both aimed at compressing LLMs while preserving their advanced reasoning capabilities and linguistic diversity. We first examine key methodologies in KD, such as task-specific alignment, rationale-based training, and multi-teacher frameworks, alongside DD techniques that synthesize compact, high-impact datasets through optimization-based gradient matching, latent space regularization, and generative synthesis. Building on these foundations, we explore how integrating KD and DD can produce more effective and scalable compression strategies. Together, these approaches address persistent challenges in model scalability, architectural heterogeneity, and the preservation of emergent LLM abilities. We further highlight applications across domains such as healthcare and education, where distillation enables efficient deployment without sacrificing performance. Despite substantial progress, open challenges remain in preserving emergent reasoning and linguistic diversity, enabling efficient adaptation to continually evolving teacher models and datasets, and establishing comprehensive evaluation protocols. By synthesizing methodological innovations, theoretical foundations, and practical insights, our survey charts a path toward sustainable, resource-efficient LLMs through the tighter integration of KD and DD principles.

Fig. 2

Overview of knowledge distillation in LLMs. Knowledge is distilled from a teacher LLM, which has been trained on a large existing database. This knowledge, potentially enriched with current, task-specific data, is transferred to a smaller student LLM. By learning from both the teacher’s guidance and the current data, the student LLM becomes more efficient and effective at performing downstream tasks

Human attention, such as where programmers look while reading or writing code, provides valuable signals that are not yet fully used in training large language models (LLMs) for code. These signals offer insights that go beyond machine-driven attention. However, collecting eye-tracking data is complex and costly, and there has been little progress in systematically applying these signals for training code LLMs.

To address this, we propose a full pipeline that combines data augmentation and reward-based fine-tuning. Specifically, we introduce: (1) an eye-tracking path augmentation method to expand programmer attention datasets, (2) a pattern abstraction step that transforms raw fixations into learnable attention motifs, and (3) a reward-guided strategy that integrates these insights into a CodeT5 supervised fine-tuning process.

Our experiments show a +7.16 improvement in CodeBLEU on the CodeXGlue benchmark for code summarization, demonstrating that combining human and machine attention can significantly enhance code intelligence. We hope this work encourages further exploration of human-centered approaches in next-generation AI for Software Engineering (AI4SE).

This study presents a new hybrid method to identify and analyze techniques used to study microplastics. By combining pattern-recognition software (regex) with the Llama 3.2:3b language model, we can better detect and understand both traditional and emerging techniques. Established methods like Raman and FTIR spectroscopy are examined alongside advanced tools such as X-ray Photoelectron Spectroscopy (XPS) and Surface-Enhanced Raman Spectroscopy (SERS). This approach improves both the speed and accuracy of identifying complex terms used in microplastics research. Using VOSDataAnalyzer and VOSviewer, we mapped connections and trends among related terms, identifying the 15 most commonly used and emerging techniques. Our analysis shows a shift toward more sensitive and innovative methods in microplastic studies. This Regex-Llama approach, introduced here for the first time, can be applied broadly to tasks such as studying pollutants in the environment, evaluating material breakdown in engineering, and assessing the health impacts of tiny contaminants. Overall, this strategy helps support environmental assessments and guide pollution reduction efforts across multiple fields.

Figure 1. �� Representation of the chemical structure of the most common polymers found in microplastic pollution, in sequence: Polyethylene (PE), Polypropylene (PP), Polystyrene (PS), and Polyethylene Terephthalate (PET).

]]>