This position paper highlights a growing trend in Explainable AI (XAI) research where Large Language Models (LLMs) are used to translate outputs from explainability techniques, like feature-attribution weights, into a natural language explanation. While this approach may improve accessibility or readability for users, recent findings suggest that translating into human-like explanations does not necessarily enhance user understanding and may instead lead to overreliance on AI systems. When LLMs summarize XAI outputs without surfacing model limitations, uncertainties, or inconsistencies, they risk reinforcing the illusion of interpretability rather than fostering meaningful transparency. We argue that - instead of merely translating XAI outputs - LLMs should serve as constructive agitators, or devil's advocates, whose role is to actively interrogate AI explanations by presenting alternative interpretations, potential biases, training data limitations, and cases where the model's reasoning may break down. In this role, LLMs can facilitate users in engaging critically with AI systems and generated explanations, with the potential to reduce overreliance caused by misinterpreted or specious explanations.

High-stakes domains like cyber operations need responsible and trustworthy AI methods. While large language models (LLMs) are becoming increasingly popular in these domains, they still suffer from hallucinations. This research paper provides learning outcomes from a case study with LinkQ, an open-source natural language interface that was developed to combat hallucinations by forcing an LLM to query a knowledge graph (KG) for ground-truth data during question-answering (QA). We conduct a quantitative evaluation of LinkQ using a well-known KGQA dataset, showing that the system outperforms GPT-4 but still struggles with certain question categories - suggesting that alternative query construction strategies will need to be investigated in future LLM querying systems. We discuss a qualitative study of LinkQ with two domain experts using a real-world cybersecurity KG, outlining these experts' feedback, suggestions, perceived limitations, and future opportunities for systems like LinkQ.

Large Language Models (LLMs) are increasingly used to automate relevance judgments for information retrieval (IR) tasks, often demonstrating agreement with human labels that approaches inter-human agreement. To assess the robustness and reliability of LLM-based relevance judgments, we systematically investigate impact of prompt sensitivity on the task. We collected prompts for relevance assessment from 15 human experts and 15 LLMs across three tasks~ -- ~binary, graded, and pairwise~ -- ~yielding 90 prompts in total. After filtering out unusable prompts from three humans and three LLMs, we employed the remaining 72 prompts with three different LLMs as judges to label document/query pairs from two TREC Deep Learning Datasets (2020 and 2021). We compare LLM-generated labels with TREC official human labels using Cohen's $κ$ and pairwise agreement measures. In addition to investigating the impact of prompt variations on agreement with human labels, we compare human- and LLM-generated prompts and analyze differences among different LLMs as judges. We also compare human- and LLM-generated prompts with the standard UMBRELA prompt used for relevance assessment by Bing and TREC 2024 Retrieval Augmented Generation (RAG) Track. To support future research in LLM-based evaluation, we release all data and prompts at https://github.com/Narabzad/prompt-sensitivity-relevance-judgements/.

A BSTRACT In this work, we present a benchmark that consists of Jupyter notebooks development trajectories and allows measuring how large language models (LLMs) can leverage runtime information for predicting code output and code generation. We demonstrate that the current generation of LLMs performs poorly on these tasks and argue that there exists a significantly understudied domain in the development of code-based models, which involves incorporating the runtime context. 1 I NTRODUCTION Recent developments in code completion and generation have been significant. Over the past several years, the field has progressed from generating relatively simple programs (Chen et al., 2021) to solving real-world issues within software repositories (Jimenez et al., 2023). However, most studies in this area are based on static snapshots of code (Jiang et al., 2024), with only a small body of research exploring the potential of leveraging dynamic code properties, such as runtime information and memory state, for code generation (Chen et al., 2024). A key reason for this limitation is that common programming environments rarely allow code generation during execution, which is when runtime information can be gathered.

Mixture-of-Experts based large language models (MoE LLMs) have shown significant promise in multitask adaptability by dynamically routing inputs to specialized experts. Despite their success, the collaborative mechanisms among experts are still not well understood, limiting both the interpretability and optimization of these models. In this paper, we focus on two critical issues: (1) identifying expert collaboration patterns, and (2) optimizing MoE LLMs through expert pruning. To address the first issue, we propose a hierarchical sparse dictionary learning (HSDL) method that uncovers the collaboration patterns among experts. For the second issue, we introduce the Contribution-Aware Expert Pruning (CAEP) algorithm, which effectively prunes low-contribution experts. Our extensive experiments demonstrate that expert collaboration patterns are closely linked to specific input types and exhibit semantic significance across various tasks. Moreover, pruning experiments show that our approach improves overall performance by 2.5\% on average, outperforming existing methods. These findings offer valuable insights into enhancing the efficiency and interpretability of MoE LLMs, offering a clearer understanding of expert interactions and improving model optimization.

Validating Large Language Models with ReLM explores the application of formal languages to evaluate and control Large Language Models (LLMs) for memorization, bias, and zero-shot performance. Current approaches for evaluating these types behavior are often slow, imprecise, costly, or introduce biases of their own, but are necessary due to the importance of this behavior when productionizing LLMs. This project reproduces key results from the original ReLM paper and expounds on the approach and applications with an emphasis on the relevance to the field of systems for machine learning.

There is a long history of building predictive models in healthcare using tabular data from electronic medical records. However, these models fail to extract the information found in unstructured clinical notes, which document diagnosis, treatment, progress, medications, and care plans. In this study, we investigate how answers generated by GPT-4o-mini (ChatGPT) to simple clinical questions about patients, when given access to the patient's discharge summary, can support patient-level mortality prediction. Using data from 14,011 first-time admissions to the Coronary Care or Cardiovascular Intensive Care Units in the MIMIC-IV Note dataset, we implement a transparent framework that uses GPT responses as input features in logistic regression models. Our findings demonstrate that GPT-based models alone can outperform models trained on standard tabular data, and that combining both sources of information yields even greater predictive power, increasing AUC by an average of 5.1 percentage points and increasing positive predictive value by 29.9 percent for the highest-risk decile. These results highlight the value of integrating large language models (LLMs) into clinical prediction tasks and underscore the broader potential for using LLMs in any domain where unstructured text data remains an underutilized resource.

Abstract: The emergence of generative AI chatbots such as ChatGPT has prompted growing public and academic interest in their role as informal mental health support tools. While early rule-based systems have been around since several years, large language models (LLMs) offer new capabilities in conversational fluency, empathy simulation, and availability. This study explores how users engage with LLMs as mental health tools by analyzing over 10,000 TikTok comments from videos referencing LLMs as mental health tools. Using a self-developed tiered coding schema and supervised classification models, we identify user experiences, attitudes, and recurring themes. Results show that nearly 20% of comments reflect personal use, with these users expressing overwhelmingly positive attitudes. Commonly cited benefits include accessibility, emotional support, and perceived therapeutic value. However, concerns around privacy, generic responses, and the lack of professional oversight remain prominent. It Is important to note that the user feedback does not indicate which therapeutic framework, if any, the LLM-generated output aligns with. While the findings underscore the growing relevance of AI in everyday practices, they also highlight the urgent need for clinical and ethical scrutiny in the use of AI for mental health support. This study does not endorse or encourage the use of AI tools as substitutes for professional mental health support.

Large Language Models (LLMs) are often provided as a service via an API, making it challenging for developers to detect changes in their behavior. We present an approach to monitor LLMs for changes by comparing the distributions of linguistic and psycholinguistic features of generated text. Our method uses a statistical test to determine whether the distributions of features from two samples of text are equivalent, allowing developers to identify when an LLM has changed. We demonstrate the effectiveness of our approach using five OpenAI completion models and Meta's Llama 3 70B chat model. Our results show that simple text features coupled with a statistical test can distinguish between language models. We also explore the use of our approach to detect prompt injection attacks. Our work enables frequent LLM change monitoring and avoids computationally expensive benchmark evaluations.

Large language models (LLMs) face significant challenges in specialized biomedical tasks due to the inherent complexity of medical reasoning and the sensitive nature of clinical data. Existing LLMs often struggle with intricate medical terminology and the need for accurate clinical insights, leading to performance reduction when quantized for resource-constrained deployment. To address these issues, we propose Quantized Medical Tree of Thought (QM-ToT), a path-based reasoning framework. QM-ToT leverages a Tree of Thought (ToT) reasoning approach to decompose complex medical problems into manageable subtasks, coupled with evaluator assessment layers. This framework facilitates substantial performance improvements in INT4-quantized models on the challenging MedQAUSMLE dataset. Specifically, we demonstrate a remarkable accuracy increase from 34% to 50% for the LLaMA2-70b model and from 58.77% to 69.49% for LLaMA-3.1-8b. Besides, we also proposed an effect data distillation method based on ToT. Compared to the traditional distillation method, we achieved an improvement of 86. 27% while using only 3.9% of the data.This work, for the first time, showcases the potential of ToT to significantly enhance performance on complex biomedical tasks, establishing a crucial foundation for future advances in deploying high-performing quantized LLM in resource-limited medical settings.