ChatDBG: Augmenting Debugging with Large Language Models

JupyterLab is a web-based interactive development platform that is widely used in the Earth science community. Using Jupyter Notebooks, it is possible to perform data analysis tasks, annotate and visualize results in a way that is easy to reproduce, present and share with others. JupyterLab allows the use of “extensions”, which add functionality to the platform. One of these is Jupyter-AI [1], which allows the use of Large Language Models (LLMs), such as ChatGPT, Claude Sonnet and Ollama, within the JupyterLab environment, through  a chat interface or directly within notebooks. By integrating LLMs into JupyterLab, it is possible to leverage their code generation capabilities to assist a user to translate their analysis tasks from an idea to actual executable code in an efficient manner. One drawback of using these LLMs in tasks involving spatio–temporal data is that the models typically do not have access to the data necessary for the analysis task and will often resort to generating fictional data or using placeholders in the code that they create. This requires the user to adapt the provided code to their data, which removes some of the utility provided by the LLM.In this context we make use of FrevaGPT, an approach for using LLMs in climate data analysis that allows for quick, complex and reproducible analyses of data sets, such as decadal climate model forecasts. Leveraging LLM’s capability to write code and using few-shot prompting (in-context learning) allows the LLM to utilize Freva [2,3] (Free Evaluation Framework), a data search and analysis platform, which provides a standardised interface to spatio-temporal datasets hosted on an HPC cluster [4].  FrevaGPT integrates seamlessly into Jupyter-AI and, by making use of the Freva library, combines the code-generating capabilities of LLMs with contextual understanding of how to access relevant datasets on the HPC cluster. This in addition with FrevaGPT’s ability to execute generated code in an isolated environment on an HPC node, annotating and explaining any intermediate results, as well as automatically correcting errors encountered along the way, could serve as a starting ramp for researchers to efficiently produce new analysis products based on spatio-temporal climate data. This PICO will include examples of using FrevaGPT within JupyterLab to analyse spatio-temporal datasets from the climate of the past, as well seasonal to decadal climate predictions.  

Engineering projects in desertification-affected regions like Valparaso, Chile, must address complex environmental, technical, and socio-economic challenges such as water scarcity and soil degradation. Traditional Root Cause Analysis (RCA) methods often fall short in these dynamic contexts due to limited scalability and adaptability. This study presents a novel methodology integrating Artificial Intelligence (AI), Large Language Models (LLMs), and Automatic Speech Recognition (ASR) to enhance RCA in environmental adaptation and mitigation efforts. The approach leverages the 9 Windows Model from the TRIZ methodology for multi-level, time-scaled problem analysis. It involves three stages: (1) collecting and transcribing environmental discussions via ASR, (2) using LLMs to extract RCA insights aligned with the 9 Windows framework, and (3) generating automated reports with visualizations and strategic recommendations. A case study in Valparaso examines the impact of desertification on water availability and agricultural productivity, demonstrating improved decision-making speed and quality. The approach reduces diagnostic time and supports more effective mitigation strategies. While AI-related challenges like bias and data dependency persist, the study emphasizes the importance of a human-in-the-loop model. This research offers a scalable, structured framework for applying AI to environmental management and supports innovation in multidisciplinary problemsolving.

This chapter introduces essential systems and tools for leveraging various natural language processing (NLP) tasks and large language models (LLMs) in qualitative analysis, emphasizing accessibility for noncoders. It begins with an overview of the Python programming language and introduces tools such as the Pandas library, integrated development environments (IDEs), Jupyter notebooks, and application programming interfaces (APIs) for accessing LLMs. A second section provides a hands-on guide for setting up a computer environment, with step-by-step code examples designed to familiarize readers with NLP libraries and LLMs. It emphasizes the acquisition of programming skills to better adapt and utilize LLM capabilities for qualitative data analysis while providing multiple resources for further learning.

To address the challenges commonly encountered in civil aircraft operating under multi-mode, strongly coupled closed-loop control—namely scarce fault samples, pronounced distribution shift, and root-cause explanations that are easily confounded by covariates—this paper proposes a root-cause analysis method that integrates ensemble learning with constraint-guided reasoning by large language models (LLMs). First, for Full Authority Digital Engine Control (FADEC) monitoring sequences, a feature system comprising environment-normalized ratios, mechanism-informed mixing indices, and multi-scale temporal statistics is constructed, thereby improving cross-mode comparability and enhancing engineering-semantic expressiveness. Second, in the anomaly detection stage, a cost-sensitive LightGBM model is adopted and a validation-set-based adaptive thresholding strategy is introduced to achieve robust identification under highly imbalanced fault conditions. Furthermore, for Root Cause Analysis (RCA), a “computation–reasoning decoupling” framework is developed: Shapley Additive exPlanations (SHAP) are used to generate segment-level contribution evidence, while causal chains, engineering prohibitions, and structured output templates are injected into prompts to constrain the LLM, enabling it to infer root-cause candidates and produce structured explanations under mechanism-consistency constraints. Experiments on real flight data demonstrate that our method yields an anomaly detection F1-score of 0.9577 and improves overall RCA accuracy to 97.1% (versus 62.3% for a pure SHAP baseline). Practically, by translating complex high-dimensional data into actionable natural language diagnostic reports, the proposed method provides reliable and interpretable decision support for rapid RCA.

The Theory of Constraints was used as a method in a root cause analysis (RCA) to identify system performance bottlenecks, wherein the constraint with the strongest correlation with system performance is identified as the bottleneck. In the present study, this method was applied in a large district cooling plant to diagnose steam absorption chillers’ performance anomalies including performance deterioration, spikes, large variation, etc. Many root causes were identified through a ten months’ continuous diagnosis, such as drastic reduction in chilled water flow rate, faulty steam control valve, cooling demand reduction, etc. More than 30 anomalies were diagnosed and more than 80% diagnosis accuracy was achieved. The diagnosis results were verified by the plant operator, which proved the effectiveness of the Theory of Constraints as a method for root cause analysis in diagnosing anomalies in district cooling plant operations. It was however discovered that even though the operational component identification accuracy was high, the determination of the actual root cause still presents lots of challenges, due to the large set of domain knowledge and operation conditions required. A knowledge-based supporting tool may be needed as part of the RCA to facilitate a speedy and complete analysis to enable these anomalies to be resolved.

Background Thoroughly investigating single patient result errors may identify systemic, large-scale testing errors. The laboratory observed an event where a revision to a Total Bile Acid (TBA, Diazyme Laboratories, Inc.) patient result following a 22S quality control (QC) failure displayed a larger than expected difference (52 to 5mcmol/L). TBA is primarily used for diagnosis and monitoring of intrahepatic cholestasis of pregnancy with results >10mcmol/L considered elevated. Retrospective review of result revisions following QC failures (3/15/2017-5/30/2023) revealed that 50%(20/40) yielded result differences of >10mcmol/L(>20SD based on assay imprecision). The aim of this study was to apply a systematic approach to i) estimate the rate of TBA errors, ii) ensure accurate result reporting during investigation period, iii) perform root cause analysis (RCA), and iv) determine corrective and preventative action(s). Methods Residual samples from TBA testing (6/24-7/5/2023) were retested(n=158), differences >+/-3.0mcmol/L or 15% were confirmed on an alternate analyzer, and reports were revised. Automated repeat testing of TBA samples >10mcmol/L was operationalized 7/6/2023. Initial and repeat TBA results were compared(n=448) and results differing by >+/-20% were remeasured on an alternate analyzer. RCA was conducted using a fishbone diagram. NaOH reagent probe washes were implemented for TBA and the error rate was re-assessed. Assays run prior to TBA samples with errors (n=15) were identified. To assess reagent carryover a residual serum pool (TBA ∼5mcmol/L) was aliquoted into 5 tubes in a sample rack. The first sample in the rack was programmed to run an assay suspected of causing carryover followed by 4 TBA measurements. Mean±standard deviation(SD) TBA concentrations were calculated. TBA was also measured in the liquid reagent for amylase, lipase, acetylcholinesterase (ACE) and fructosamine. Results Initial TBA retesting yielded no errors when initial TBA was ≤10mcmol/L(n=51). For samples with TBA >10mcmol/L(n=107), 9(8.4%) had differences exceeding criteria with 8/9 being revised to ≤10mcmol/L. Analysis of automated patient repeat data showed a 3.8%(17/448) error rate when initial TBA >10mcmol/L. After NaOH reagent probe washes were implemented, the error rate decreased to 0.7%(3/448). Assays run directly before an erroneously high TBA result included: lactate, fructosamine, soluble transferrin receptor(STFR), lipase, and ACE. A serum TBA pool(mean±SD=5.1±0.4 mcmol/L,n=15) measured 37.6±1.2mcmol/L(n=3) and 6.7±0.7mcmol/L(n=3) after lipase and fructosamine, respectively. No other assays demonstrated carryover. TBA in lipase reagent compartment B and C was 653 and 649mcmol/L, respectively, and 653mcmol/L in fructosamine reagent compartment B. Conclusions A single patient TBA result revision was investigated and led to identification of reagent carryover causing erroneously high TBA results on the Roche c502 analyzer. Automatic, real-time, repeat testing was implemented to prevent reporting incorrect patient results until RCA could identify the cause. Occurrence rate of erroneous high TBA results decreased from 3.8% to 0.7% after implementation of NaOH washes on the instruments. Carryover experiments confirmed that lipase and fructosamine assays cause carryover when run prior to a TBA sample due to measurable TBA in the reagent. The lab is pursuing moving the fructosamine and lipase reagent to the Roche Cobas c701 to prevent the issue.

Collaborative Problem Solving (CPS) is a vital 21st-century skill that demands nuanced coordination of cognitive, social, and regulatory processes among learners. In face-to-face classrooms, peer dialogue offers rich data for studying CPS, but manual annotation of such unstructured, oral interaction is labour-intensive and difficult to scale. This study investigates whether open-source Large Language Models (LLMs), including Llama and Qwen, can perform inductive qualitative coding on classroom peer dialogues using zero-shot prompting alone—without fine-tuning or training data. We collected over 210,000 words of dialogue transcripts from 38 student dyads across two CPS tasks at university classrooms: an engineering design activity and a GenAI-supported peer assessment of lesson plans. Through a multi-phase process, we iteratively developed three zero-shot prompting strategies (self-prompting, chain-of-thought prompting, and in-context prompting) via GPT-4o interactions and deployed them across different LLMs via API access. Our findings suggest that in-context prompting consistently yields context-sensitive and theoretically coherent CPS constructs. Chain-of-thought prompting facilitates abstract reasoning but may lead to overgeneralization, while self-prompting demonstrates autonomous logic yet lacks consistency. Expert evaluations using a five-dimensional rubric (clarity, concreteness, objectivity, granularity, specificity) show moderate to high alignment between human and LLM-generated codes, although LLMs tend to overrate clarity and coherence. We further analyse discrepancies between LLMs and academic frameworks such as PISA CPS and ATC21S, and highlight challenges such as terminological drift, low recall, and theoretical misalignment. This work contributes a scalable, human-centered workflow for inductive coding of classroom dialogue and provides ready-to-use prompt templates for educational researchers. The aim of this study is to critically examine whether open-source LLMs can inductively code classroom peer dialogue in collaborative problem-solving tasks, while acknowledging both their potential and limitations in educational practice. We conclude with a dual-pathway strategy for combining practice-oriented, behaviourally grounded constructs with theory-aligned coding schemes, and offer design recommendations for future human-AI collaborative tools in learning analytics and classroom assessment.

Blood and Body Fluid Exposures (BFFEs) involve a complex sequence of events combining technical, human and organizational factors. Performing root cause analysis (RCA) of these events is promoted to improve safety. We conducted a RCA of a BBFE in a radiology unit using a recently developed French method called Orion.

UAV accident investigation is essential for safeguarding the fast-growing low-altitude airspace. While near-daily incidents are reported, they were rarely analyzed in depth as current inquiries remain expert-dependent and time-consuming. Because most jurisdictions mandate formal reporting only for serious injury or substantial property damage, a large proportion of minor occurrences receive no systematic investigation, resulting in persistent data gaps and hindering proactive risk management. This study explores the potential of using large language models (LLMs) to expedite UAV accident investigations by extracting human-factor insights from unstructured narrative incident reports. Despite their promise, the off-the-shelf LLMs still struggle with domain-specific reasoning in the UAV context. To address this, we developed a human factors analysis and classification system (HFACS)-guided analytical framework, which blends structured prompting with lightweight post-processing. This framework systematically guides the model through a two-stage procedure to infer operators’ unsafe acts, their latent preconditions, and the associated organizational influences and regulatory risk factors. A HFACS-labelled UAV accident corpus comprising 200 abnormal event reports with 3600 coded instances has been compiled to support evaluation. Across seven LLMs and 18 HFACS categories, macro-F1 ranged 0.58–0.76; our best configuration achieved macro-F1 0.76 (precision 0.71, recall 0.82), with representative category accuracies > 93%. Comparative assessments indicate that the prompted LLM can match, and in certain tasks surpass, human experts. The findings highlight the promise of automated human factor analysis for conducting rapid and systematic UAV accident investigations.

The gut microbiome plays a fundamental role in host metabolism, immune regulation, and disease development. With the rapid accumulation of multi-omics and literature data, the microbiome field now faces the challenge of efficiently extracting scientific insights from massive, heterogeneous datasets. Artificial intelligence (AI) and large language models (LLMs) provide promising tools to address this complexity by enabling integrative analysis and knowledge synthesis across diverse biological sources. In this study, we developed METABOLISM, a microbiome-specialized LLM fine-tuned on 160,000 scientific abstracts to enhance literature-based contextual understanding of microbiome-liver interactions and related biological mechanisms. Using LoRA-based parameter-efficient training, METABOLISM was optimized for domain-specific reasoning and response generation. Model performance was evaluated through both automated Phi-4 scoring (a large language model-based evaluator for relevance, informativeness, and fluency) and structured human expert rubric assessments involving 20 domain specialists. The fine-tuned METABOLISM achieved superior relevance and clarity scores (mean > 7.5 ± 0.06) compared with general-purpose LLMs such as Gemma-3-12B-IT and ChatGPT-4o. Correlation analysis revealed weak to moderate negative relationships (R = -0.65, p < 0.0001) between traditional NLP metrics (BLEU, ROUGE) and human expert rubric scores, with a similar trend observed for correlations with Phi-4-based automated evaluation scores, indicating the limitations of surface-level similarity measures in biomedical contexts. Overall, our findings demonstrate that microbiome-adapted LLMs can effectively distill high-volume scientific data into biologically meaningful insights, supporting more efficient and interpretable research in microbiology and systems biology.

Research Question: This paper investigates into how machine learning can be applied for the purpose of detecting anomalies in the data describing transport component within the cellular network. Motivation: In the field of telecommunications, terabytes of data are generated each hour. This makes the manual analysis almost impossible to perform. There are thousands of components whose behaviour needs to be monitored, since anomalous behaviour could indicate a possible failure that can lead to network degradation, huge maintenance costs, and finally – a bad user experience. Our goal is to try to catch anomalous behaviour automatically, and thus help domain experts when performing drill down analysis of the degraded base stations and their key performance indicators (KPIs). Idea: The main idea of this paper is to empirically evaluate the application of machine learning for the problem of anomaly detection, in the field of telecommunications, specifically to long term evolution (LTE) networks. Data: Data used in the analysis contains information about base transceiver stations (BTS) behaviour through the time. The data are gathered from a cellular network provider located in Serbia. The data are collected on an hourly basis, for a period of two weeks, resulting in almost 700 thousand rows. The behaviour is assessed by 96 transport KPIs coming from BTS, describing the package losses, delays, transmission success rates, etc. Tools: Two main algorithms, ensemble-based Isolation Forest and autoencoder neural network, are elaborated and applied in order to identify patterns of anomalous behaviour. Findings: The results show that machine learning can be successfully applied in the field of LTE networks for the problem of anomaly detection. Machine learning can significantly reduce the time needed for the domain experts to identify anomalies within the network. In addition to time efficiency, one of the algorithms tested is able to identify anomalous KPIs separately, which is crucial when performing root cause analysis, by using drill-down approach, in order to identify which component is degraded. Contribution: This paper enriches existing research related to anomaly detection in LTE networks and provides an innovative approach to automated root-cause analysis of network degradation.

BackgroundThe growing demand for genomic testing and limited access to experts necessitate innovative service models. While chatbots have shown promise in supporting genomic services like pre-test counseling, their use in returning positive genetic results, especially using the more recent large language models (LLMs) remains unexplored.ObjectiveThis study reports the prompt engineering process and intrinsic evaluation of the LLM component of a chatbot designed to support returning positive population-wide genomic screening results.MethodsWe used a three-step prompt engineering process, including Retrieval-Augmented Generation (RAG) and few-shot techniques to develop an open-response chatbot. This was then evaluated using two hypothetical scenarios, with experts rating its performance using a 5-point Likert scale across eight criteria: tone, clarity, program accuracy, domain accuracy, robustness, efficiency, boundaries, and usability.ResultsThe chatbot achieved an overall score of 3.88 out of 5 across all criteria and scenarios. The highest ratings were in Tone (4.25), Usability (4.25), and Boundary management (4.0), followed by Efficiency (3.88), Clarity and Robustness (3.81), and Domain Accuracy (3.63). The lowest-rated criterion was Program Accuracy, which scored 3.25.DiscussionThe LLM handled open-ended queries and maintained boundaries, while the lower Program Accuracy rating indicates areas for improvement. Future work will focus on refining prompts, expanding evaluations, and exploring optimal hybrid chatbot designs that integrate LLM components with rule-based chatbot components to enhance genomic service delivery.

Root Cause Analysis (RCA) of reported incident reports can lead to measures to prevent the recurrence of accidents. The purpose of this study is to clarify the relationship between the occurrence of cases, causal factors, contributing factors, and the experience years of the reporters classified into three groups: less than 2 years, 3 to less than 5 years, and more than 5 years, for incident reports subject to RCA. From April 2018 to March 2023, a chi-square test was conducted between each item extracted from 239 cases subject to RCA and the experience years of the reporters, with a significance level of less than 5% considered significant. Regarding the occurrence cases, radiological technologists with longer experience had more misidentified a patient and fewer errors in imaging conditions and range, while radiological technologists with 3 to less than 5 years of experience had more errors in imaging conditions (p<0.001). In terms of occurrence factors, radiological technologists with 3 to less than 5 years of experience had more cases of insufficient confirmation (p<0.05). For contributing factors, there was no significant difference between the experience year groups (p=0.19), with "Impatience" being the most common factor. This survey suggested that in incident reports of radiological technologists, "Impatience" is a contributing factor that can occur regardless of years of experience.

Building operations still rely heavily on manual review of unstructured maintenance records, which is slow, error-prone, and difficult to standardize across systems and teams. We present a knowledge-graph-centric middleware framework that mediates between heterogeneous IT/OT sources and analytics applications to support reliable, low-latency root-cause analysis. The framework integrates large language models (LLMs) with a domain knowledge graph (KG) under explicit quality criteria to ensure reliable construction and evolution. It begins with LLM-assisted schema generation, guided by Knowledge Graph Schema (KGS) metrics that ensure consistency, completeness, scalability, and minimal redundancy. Building on this foundation, the graph structure is refined using Graph Convolutional Network (GCN) embeddings combined with recursive KMeans clustering, which groups similar types and lowers query overhead. Next, a two-stage extraction pipeline converts unstructured text into a structured knowledge data frame and enriches relations by leveraging contextual proximity through weighted aggregation. To accommodate evolving data, the system supports incremental updates for both seen and unseen entities, preserving chronological consistency while avoiding disruptive schema revisions. Together, these components allow applications to remain decoupled from heterogeneous source formats while ensuring robust governance of both schema and data. The design further reduces query latency, improves retrieval precision, and reduces manual effort, thus supporting more timely and consistent maintenance decisions.

Large language models (LLMs) have shown great promise in automating data science workflows. However, existing models still struggle with multi-step reasoning and tool use, limiting their effectiveness on complex data analysis tasks. To address this limitation, we propose a scalable pipeline that extracts high-quality, tool-based data analysis tasks and their executable multi-step solutions from real-world Jupyter notebooks and associated data files. Using this pipeline, we introduce NbQA, a large-scale dataset of standardized task–solution pairs that reflect authentic tool-use patterns in practical data science scenarios. To further enhance the multi-step reasoning capabilities, we present Jupiter, a framework that formulates data analysis as a search problem and applies Monte Carlo Tree Search (MCTS) to generate diverse solution trajectories for value model learning. During inference, Jupiter combines the value model and node visit counts to efficiently collect executable multi-step plans with minimal search steps. Experimental results show that Qwen2.5-7B and 14B-Instruct models on NbQA solve 77.82% and 86.38% of tasks on InfiAgent-DABench, respectively—matching or surpassing GPT-4o and advanced agent frameworks. Further evaluations demonstrate improved generalization and stronger tool-use reasoning across diverse multi-step reasoning tasks.