Large language models (LLMs) are being increasingly used for mental health and informational support, yet little is known about their effectiveness in addressing problem gambling inquiries. This study evaluates how two general-purpose LLMs—OpenAI’s GPT-4o and Meta’s Llama 3.1 (405B)—respond to nine problem gambling questions constructed based on the Problem Gambling Severity Index (PGSI). We collected responses by prompting each LLM via its respective chatbot interface (ChatGPT-4o and Meta AI) and recruited professional gambling counselors (n = 23) to provide their own responses via an online survey. We asked counselors which chatbot responses they preferred and whether exposure to chatbot responses influenced their willingness to alter their answers. We compared LLM and human responses by analyzing several linguistic and readability metrics. Our results reveal that LLMs generate more verbose responses and that counselors prefer Llama’s responses over GPT’s. Most counselors reported that they would not change their own responses after reviewing the LLM-generated responses. This preliminary comparison highlights important considerations for integrating LLM-based tools into gambling harm prevention strategies by comparing their responses to those of experts with extensive experience in providing in-person treatment for gambling problems.

Large language models (LLMs), especially those capable of retrieval-augmented generation (RAG), can be used to answer questions such as, “What is the best treatment for a migraine?” Here, a good answer depends on the specific intent (best treatment for what goal?) and the interpretation of the question (best treatment according to whom?). We present CORAL, an interactive tool for exploring answers to open-ended questions generated by RAG LLMs in simple terms. To achieve this, we structure the answer space as a lattice of semantic concepts representing different intents and interpretations. Using this formulation, we demonstrate how CORAL allows users to efficiently explore this space, with applications to downstream tasks such as query refinement and LLM safeguarding.

Modern database systems including IBM Db2 rely on sophisticated cost-based optimizers, often augmented with machine learning (ML), to generate efficient query execution plans (QEPs) [1]–[6]. However, these types of ML “black box” systems incorporate complex algorithms, making it difficult for experts to diagnose performance regressions and tune them [7]–[9]. Traditional eXplainable AI (XAI) techniques, such as SHAP and LIME, assign importance scores by perturbing model inputs [10]–[12]. However, these methods assume well-defined perturbation strategies, such as removing or masking features in text [13] or images [14], which are invalid for structured SQL queries whose semantics must remain intact.

Modern data-centric models, such as large language models (LLMs), are difficult to understand in part due to their unstructured interface. For example, LLMs using retrieval-augmented generation (RAG) condition on external knowledge sources, but it is unclear how these sources shape their reasoning and responses [4], [10]. Without structured explanations of their output, LLMs are difficult to control and debug, ultimately compromising user trust and safety [2], [3]. To advance LLM explainability, we seek to reintroduce structure by mapping LLM inputs and outputs back to interpretable structures. This structure enables users to reason about model behavior, facilitating explainability, provenance, and other insights necessary to build trust.

Online communities have become essential places for socialization and support, yet they also possess toxicity, echo chambers, and misinformation. Detecting this harmful content is difficult because the meaning of an online interaction stems from both what is written (textual content) and where it is posted (social norms). We propose GASTON (Graph-Aware Social Transformer for Online Networks), which learns text and user embeddings that are grounded in their local norms, providing the necessary context for downstream tasks. The heart of our solution is a contrastive initialization strategy that pretrains community embeddings based on user membership patterns, capturing a community's user base before processing any text. This allows GASTON to distinguish between communities (e.g., a support group vs. a hate group) based on who interacts there, even if they share similar vocabulary. Experiments on tasks such as stress detection, toxicity scoring, and norm violation demonstrate that the embeddings produced by GASTON outperform state-of-the-art baselines.

We present HEX, a tool for generating simple and personalized explanations of machine learning model predictions. HEX draws inspiration from Online Analytical Processing (OLAP) and data cubes, treating model features as hierarchical dimension attributes and using roll-up and drill-down operations to aggregate explanations at varying levels of detail. We also develop a tunable informativeness metric that allows users to balance explanatory insight and simplicity by selectively expanding the most informative parts of the hierarchy. Finally, we demonstrate HEX's utility in interpreting model decisions in high-dimensional feature spaces and summarizing feature perturbations to identify model-decision boundaries.

Model slicing refers to computing model-performance metrics independently for sub-groups to identify biases. As an example, a healthcare model with good overall accuracy may perform poorly for children or seniors. We extend model slicing to summarize a new type of bias: algorithmic recourse. This quantifies the ability and cost, on average, for members of sub-groups to move from an undesirable to a desirable outcome. As an example, a recourse-biased loan-approval model may show equal classification accuracy for different sub-groups; however, women whose loans were declined may need to increase their savings balance by twenty percent to receive approval, while men only require a ten-percent increase. Our solution, REACT, combines counterfactual mining, finding nearest positively labeled neighbors to a given negatively labeled example, and if-then rule mining to generate a concise summary as an explanation table of recourse bias across sub-groups. We experimentally demonstrate the value of our solution using several case studies including finance and law enforcement.

If-then rules are widely used to explain machine learning models; e.g., "if employed = no, then loan application = rejected." We present the first proposal to apply rules to explain the emerging class of large language models (LLMs) with retrieval-augmented generation (RAG). Since RAG enables LLM systems to incorporate retrieved information sources at inference time, rules linking the presence or absence of sources can explain output provenance; e.g., "if a Times Higher Education ranking article is retrieved, then the LLM ranks Oxford first." To generate such rules, a brute force approach would probe the LLM with all source combinations and check if the presence or absence of any sources leads to the same output. We propose optimizations to speed up rule generation, inspired by Apriori-like pruning from frequent itemset mining but redefined within the scope of our novel problem. We conclude with qualitative and quantitative experiments demonstrating our solutions' value and efficiency.

In self-consuming generative models that train on their own outputs, alignment with user preferences becomes a recursive rather than one-time process. In this paper, we provide the first formal foundation for analyzing the long-term effects of such recursive retraining on alignment. Under a two-stage curation mechanism based on the Bradley–Terry (BT) model, we model alignment as an interaction between two factions: the Model Owner, who filters which outputs should be learned by the model, and the Public User, who determines which outputs are ultimately shared and retained through interactions with the model. Our analysis reveals three structural convergence regimes: consensus collapse, compromise on shared optima, and asymmetric refinement, depending on the degree of preference alignment. We prove a fundamental impossibility theorem: no recursive BT-based curation mechanism can simultaneously preserve diversity, ensure symmetric influence, and eliminate dependence on initialization. Framing the process as dynamic social choice, we show that alignment is not a static goal but an evolving equilibrium shaped by power asymmetries and path dependence.

Modern database systems including IBM Db2 rely on cost-based optimizers to select efficient query execution plans (QEPs), but these complex “black box” mechanisms make performance tuning challenging. We present GeX, an eXplainable AI (XAI) system that builds surrogate models to approximate optimizer behavior and generate query and QEP saliency maps. These maps highlight predicates and operators most influential to cost estimates, guiding experts toward tuning actions, including statistics collection, knob adjustment, and query rewrites. We demonstrate how GeX assists users in diagnosing performance issues, supports tuning in downstream tasks and improves workload runtimes beyond expert-crafted strategies.

Model slicing is a machine learning diagnostic method that partitions data into coherent subgroups (slices), such as those corresponding to protected attributes, and computes model performance metrics within each slice. As an example, model slicing might find that a healthcare model performs well overall but poorly for children and for seniors. We present a tutorial that describes model slicing from a data engineering perspective, as multi-dimensional analytics of model performance data. Within this formulation, we represent variants of model slicing in terms of their dimension and measure attributes. We also discuss performance optimizations and we conclude with directions for future work.

Large Language Models (LLMs) have demonstrated remarkable question-answering (QA) capabilities, yet their decision processes and outputs often remain opaque and prone to factual inconsistencies. While existing methods evaluate or ground LLM outputs after generation, they typically lack mechanisms for aligning LLM reasoning with external knowledge sources. This paper introduces AprèsCoT, a lightweight model-agnostic framework that validates LLM reasoning by grounding it in an external knowledge graph (KG). AprèsCoT operates through three main components: Subgraph Retrieval, which extracts a KG subgraph relevant to a given query; Triple Extraction and Parsing, which converts the LLM’s output into factual triples; and Matching, which aligns these triples with entities and relations in the extracted KG subgraph. The integration of these modules enables alignment between LLM reasoning and structured knowledge, producing traceable and structured explanations alongside model outputs. We evaluate alternative retrieval and matching strategies, analyze their trade-offs, and demonstrate how AprèsCoT helps users surface reasoning gaps, hallucinations, and missing facts. Experiments across multiple domains, including large KGs, highlight AprèsCoT’s effectiveness in advancing trustworthy and explainable AI.

Recursive retraining of generative models poses a critical representation challenge: when synthetic outputs are curated based on a fixed reward signal, the model tends to collapse onto a narrow set of outputs that over-optimize that objective, causing diversity to vanish and failing to represent the full range of preferences. Prior work has suggested that such collapse is unavoidable without adding real data into the mix. In this paper, we revisit that conclusion from an alignment perspective and show that collapse can be mitigated through curation based on multiple reward functions. We formalize the dynamics of recursive training under heterogeneous preferences and prove that, under certain conditions, the model converges to a stable distribution that allocates probability mass across competing high-reward regions. The limiting distribution preserves diversity and provably satisfies a weighted Nash bargaining solution, offering a formal interpretation of value aggregation in synthetic retraining loops.

Expert search and team formation systems operate on collaboration networks, with nodes representing individuals, labeled with their skills, and edges denoting collaboration relationships. Given a keyword query corresponding to the desired skills, these systems identify experts that best match the query. However, state-of-the-art solutions to this problem lack transparency. To address this issue, we propose ExES, a tool designed to explain expert search and team formation systems using factual and counterfactual methods from the field of explainable artificial intelligence (XAI). ExES uses factual explanations to highlight important skills and collaborations, and counterfactual explanations to suggest new skills and collaborations to increase the likelihood of being identified as an expert. Towards a practical deployment as an interactive explanation tool, we present and experimentally evaluate a suite of pruning strategies to speed up the explanation search. In many cases, our pruning strategies make ExES an order of magnitude faster than exhaustive search, while still producing concise and actionable explanations.

Large Language Models (LLMs) have transformed information retrieval for humans. People are increasingly turning to general-purpose LLM-based chatbots to find answers to questions across numerous domains, including advice on sensitive topics such as mental health and addiction. In this study, we present the first inquiry into how LLMs respond to prompts related to problem gambling, specifically exploring how experienced gambling treatment professionals interpret and reflect on these responses. We used the Problem Gambling Severity Index to develop nine prompts related to different aspects of gambling behavior. These prompts were submitted to two LLMs, GPT-4o (via ChatGPT) and Llama 3.1 405b (via Meta AI), and their responses were evaluated via an online survey distributed to human experts (experienced gambling treatment professionals). Twenty-three experts participated, representing over 17,000 hours of problem gambling treatment experience. They provided their own responses to the prompts and selected their preferred (blinded) LLM response, along with contextual feedback, which was used for qualitative analysis. Llama was slightly preferred over GPT, receiving more votes for 7 out of the 9 prompts. Thematic analysis revealed that experts identified strengths and weaknesses in LLM responses, highlighting issues such as encouragement of continued gambling, overly verbose messaging, and language that could be easily misconstrued. These findings offer a novel perspective by capturing how experienced gambling treatment professionals perceive LLM responses in the context of problem gambling, providing insights to inform future efforts to align these tools with appropriate guardrails and safety standards for use in gambling harm interventions.

Studies show that large language models (LLMs) produce buggy code translations. One avenue to improve translation accuracy is through intermediate representations, which could provide structured insights to guide the model's understanding. We explore whether code translation using LLMs can benefit from intermediate representations via natural language (NL) and abstract syntax trees (ASTs). Since prompt engineering greatly affects LLM performance, we consider several ways to integrate these representations, from one-shot to chain-of-thought (CoT) prompting. Using Open Gpt4 8X7B and specialized StarCoder and CodeGen models on popular code translation benchmarks (CodeNet and AVATAR), we find that CoT with an intermediate NL summary performs best, with an increase of 13.8

ClinicalTrials.gov hosts an online database with over 440,000 medical studies (as of 2023) evaluating drugs, supplements, medical devices, and behavioral treatments. Target users include scientists, medical researchers, pharmaceutical companies, and other public and private institutions. Although ClinicalTrials has some filtering ability, it does not provide visualization tools, reporting tools or historical data; only the most recent state of each trial is visible to users. To fill these functionality gaps, we present Tri-AL: an open-source data platform for clinical trial visualization, information extraction, historical analysis, and reporting. This paper describes the design and functionality of Tri-AL, including a programmable module to incorporate machine learning models and extract disease-specific data from unstructured trial reports, which we demonstrate using Alzheimer’s disease reporting as a case study. We also highlight the use of Tri-AL for trial participation analysis in terms of sex, gender, race and ethnicity. The source code is publicly available at https://github.com/pouyan9675/Tri-AL.