Explainable language model reasoning for laboratory quality control: validation of the ChatGPT-5 as a digital quality assistant

Speech recognition in digital assistants such as Google Assistant can potentially benefit from the use of conversational context consisting of user queries and responses from the agent. We explore the use of recurrent, Long Short-Term Memory (LSTM), neural language models (LMs) to model the conversations in a digital assistant. Our proposed methods effectively capture the context of previous utterances in a conversation without modifying the underlying LSTM architecture. We demonstrate a 4% relative improvement in recognition performance on Google Assistant queries when using the LSTM LMs to rescore recognition lattices.

The paper studies the problem of integrating language models (LM) and knowledge graph (KG). KG is built in the semantic library of scientific subject areas LibMeta for navigation through scientific publications. Using the example of KG of the mathematical subject area (SjD), it is shown that as a result of this approach, LM does not go beyond the SjD, which allows us to state a more relevant answer to the query. The descriptions of mathematical SjD are based on mathematical encyclopedias of the soviet mathematical school and the library of subject areas is filled by integrating subject areas of specialized mathematical journals. Using the example of mathematical SjD and applications, the problem of creating an environment for using a digital assistant in Russian when mastering scientific knowledge in a local SjD and accessing scientific research is considered. Setting up LM on SjD is implemented by creating a set of instructions and checking the truth of the answers based on them. Applications of the research results are expected to be implemented in mathematical knowledge systems, library and journal systems to support business processes, search and analysis of scientific publications. The research is aimed at creating a technology for information support of scientific research in the process of searching and analyzing scientific information. The proposed approach allows reducing the flow of information noise when working with scientific publications. A methodology for the interaction of LM and KG of mathematical SjD has been developed based on instructions applied to the description of SjD in the form of KG. The application of the proposed approach will allow using multiple instructions to simplify work with LM in the process of searching for specialized information while reducing LM hallucinations and without involving expert advice. In the context of intensification of scientific work associated with an increasing flow of information, a solution for search augmented generation (RAG) is proposed.

Learning curve of vitrification assessed by cumulative summation test for learning curve (LC-CUSUM)

Vocational guidance is very useful in helping students make informed academic and professional choices worldwide. However, in Peru, many young people do not have access to this type of specialized support, contributing to issues such as school dropout and poor career decision-making. To help address this gap, we developed a digital assistant software that provides vocational guidance in Spanish using the Large Language Model Meta Artificial Intelligence (LLaMA) 3.2 3B. The development process followed a four-phase methodology. First, Holland's test was selected as the psychometric tool. Second, we trained and optimized LLaMA 3.2 3B using specialized vocational guidance datasets, enabling the system to correctly interpret test responses. Third, we designed and implemented a mobile application that allows students to interact with a digital assistant via voice or text messages. Finally, we conducted a usability and effectiveness evaluation of the system with 40 students from a public high school. By comparing the assistant's recommendations to those provided by an expert psychologist, we obtained a concordance rate of 75.83%, while 80% of participating students were able to use the system without external assistance. These findings indicate that the proposed digital assistant has strong potential to serve as an effective and accessible tool for vocational guidance in Peru.

Laboratory quality control is all the measures put in place to eliminate the risk of non-conforming outcomes. It involves systems that safeguard the accuracy, reliability, and timeliness of lab results by ensuring the early detection of results or measurement errors and the procedures to rectify them. It should be performed regularly and quality control materials should be treated the same as samples, from the beginning to the end of the run. Laboratory quality control (QC) ensures that the lab processes and operations run efficiently and guarantees the production of accurate and reproducible results. In addition, the QC measures developed in a lab are the building blocks for the process of certification and accreditation. Failure to integrate quality control in a laboratory can lead to several negative consequences, including the following: Time wastage, as experiments and tests are repeated. Budget implications, as more reagents are needed to carry out repeat tests and experiments. Unreliable results, which will impact the integrity of the lab and consequently any funding options and certification/accreditation process. Loss of customer loyalty and satisfaction. Safety concerns due to non-compliance in the absence of quality control mechanisms. Delayed diagnosis or unnecessary treatments for patients. The process of setting up laboratory quality management begins with identifying all the lab processes and practices that are susceptible to inefficiencies, errors, and safety concerns, in order to build systems that secure them as discussed below.

Analytical validation is required as the basis for any evaluation activities during manufacturing process validation, cleaning validation and validation of the testing method itself in the pharmaceutical industry according to good manufacturing practice (GMP) rules and guidelines. Validation of analytical methods and procedures in a quality control (QC) laboratory is implemented mainly at the time of transfer or introduction of the methods developed by the analytical development laboratory within group companies or elsewhere. However, it is sometimes necessary to develop a new or improved method of analysis for the QC laboratory’s own use. In the first part of this report, a general description of analytical validation of the high performance liquid chromatography (HPLC) method including preparation of documents is presented based on the experience in our QC laboratory. A typical example of method validation of robotic analysis system is then cited. Finally the merits and demerits of these analytical validations for QC laboratories are summarized. The authors emphasize the importance of analytical validation and the responsibility of QC laboratory management for the effective design and implementation of validation activities.

We propose a quantum-inspired framework for modeling open distributed intelligence systems (DISs) comprising natural intelligence agents (NIAs) and artificial intelligence agents (AIAs) that interact with each other. Each NIA – AIA pair represents a user and their digital assistant – an avatar implemented as an agent based on a large language model (LLM). The AIAs are interconnected through a complex, scale-free network and communicate with users and one another in real time. We focus on the social impact and evolution of users’ emotional states, which we model as simple, two-level cognitive systems shaped by interactions with AIAs and external information sources. Within this framework, the AIAs adiabatically follow the NIAs, mediating emotional influence by disseminating information and propagating user emotions throughout the system. Building on Mehrabian’s Pleasure–Arousal–Dominance (PAD) model and Wundt’s three-dimensional theory of emotions, we put forward a quantum-like representation of affective states on an emotional sphere. We demonstrate that the arousal component is governed by the interplay between external informational inputs and individual personality traits. This leads to the emergence of limiting cycles in emotional dynamics. Assuming weak AIA – AIA coupling, we identify two distinct regimes of affective behavior. In the first regime, coherent NIA – AIA interaction supports emotional heterogeneity and individual differentiation across the network. In the second regime, shared exposure to external information drives synchronized emotional responses, resulting in a macroscopic affective field that captures collective emotional dynamics. Furthermore, we demonstrate that the network’s structural properties, particularly node degree correlations, play a role analogous to quantum correlations in ensembles of two-level physical systems; a quantum-like superradiant state corresponds to the network-induced collective emotional activation of NIAs within a DIS. These findings advance our understanding of affective dynamics and emergent social phenomena in hybrid human–AI ecosystems.

In the realm of business automation, conversational assistants are emerging as the primary method for making automation software accessible to users in various business sectors. Access to automation primarily occurs through application programming interface (APIs) and robotic process automation (RPAs). To effectively convert APIs and RPAs into chatbots on a larger scale, it is crucial to establish an automated process for generating data and training models that can recognize user intentions, identify questions for conversational slot filling, and provide recommendations for subsequent actions. In this paper, we present a technique for enhancing and generating natural language conversational artifacts from API specifications using large language models (LLMs). The goal is to utilize LLMs in the “build” phase to assist humans in creating skills for digital assistants. As a result, the system does not need to rely on LLMs during conversations with business users, leading to efficient deployment. Along with enabling digital assistants, our system employs LLMs as proxies to simulate human interaction and automatically evaluate the digital assistant's performance. Experimental results highlight the effectiveness of our proposed approach. Our system is deployed in the IBM Watson Orchestrate product for general availability.

BackgroundPhysical inactivity is prevalent, leading to a high burden of disease and large healthcare costs. Thus, there is a need for affordable, effective and scalable interventions. However, interventions that are affordable and scalable are beset with modest effects and engagement. Interventions that integrate machine learning with real-time data to offer unprecedented levels of personalisation and customisation might offer solutions. The aim of this study is to conduct a randomised controlled trial to evaluate the effectiveness of a machine learning and app-based digital assistant to increase physical activity.MethodsOne hundred and ninety-eight participants will be recruited through Facebook advertisements and randomly allocated to an intervention or control group. Intervention participants will gain access to an app-based physical activity digital assistant that can learn and adapt in real-time to achieve high levels of personalisation and user engagement by virtue of applying a range of machine learning techniques (i.e. reinforcement learning, natural language processing and large language models). The digital assistant will interact with participants in 3 main ways: (1) educational conversations about physical activity; (2) just-in-time personalised in-app notifications (‘nudges’), cues to action encouraging physical activity and (3) chat-based questions and answers about physical activity. Additionally, the app includes adaptive goal setting and an action planning tool. The control group will gain access to the intervention after the last assessment. Outcomes will be measured at baseline, 3 and 6 months. The primary outcome is device-measured (Axivity AX3) moderate-to-vigorous physical activity. Secondary outcomes include app engagement and retention, quality of life, depression, anxiety, stress, sitting time, sleep, workplace productivity, absenteeism, presenteeism and habit strength.DiscussionThe trial presents a unique opportunity to study the effectiveness of a new generation of digital interventions that use advanced machine learning methods to improve physical activity behaviour. By addressing the limitations of existing conversational agents, we aim to pave the way for more effective and adaptable interventions.Trial registrationAustralian New Zealand Clinical Trial Registry ACTRN12624000255583p. Registered on 14 March 2024. https://www.anzctr.org.au/Trial/Registration/TrialReview.aspx?id=387332.

The test methods used for PNL bioassay performance tests were evaluated by comparing the MDA based on performance tests results with MDA calculated by PNL using the bioassay laboratory's own quality control (QC) data. Two in vitro laboratories and two in vivo laboratories were studied and a correlation between the performance test MDA estimates and QC data was demonstrated. However, it was often necessary to examine the QC data to identify important characteristics of the blank distribution that affect the MDA calculation. Since the MDA equation must be based on the specific analysis and calculational methods of the procedure evaluated. Even when the correct MDA equation is applied, the MDA calculated will have a relatively large confidence interval when only a few replicates are used to estimate the standard deviation. For this reason, a relatively precise estimate of the MDA is generally only available when Poisson statistics may be applied. It was concluded that performance testing alone cannot provide all the information necessary to make an accurate estimate of the measurement process MDA. Review of the laboratory's QC data and the entire measurement procedure will be necessary. Specific recommendations for changes to draft ANSI N13.30 Performance Criteria for Radiobioassay'' are given. 10 refs., 18 figs., 11 tabs.

Objective: This study investigates the potential of Large Language Models (LLMs) to support medical prescription processes and enhance patient safety. Methods: Six LLMs answered four prescription-related questions on contraindications, drug interactions, and dosage. A panel of 34 physicians blindly evaluated 24 responses based on consistency, focus, coherence, completeness, and detail. Results: LLM performance varied by criteria and question type; LLM6 excelled in completeness and detail, especially in complex cases. Simpler questions, like contraindications, scored higher overall, while complex queries showed more variation. Conclusion: LLMs show promise as digital assistants in prescription tasks, improving access to medical info and reducing errors. However, reliability depends on question complexity. They should support, not replace, clinical judgment and require ongoing validation for healthcare adoption.

The reduction of on-site workers and the advancement of equipment lead to a lack of worker experience, resulting in human errors and decreased efficiency. To address this issue, this paper proposes a user-adapting training system using large language models. This system offers personalized assistance based on the initial background level of users. It updates the level in real time using feedback evaluated by the large language model. The proposed training system adjusts dynamically to the updated user level to provide appropriate assistance. Additionally, it identifies given materials, that need to be reinforced, through user feedback. These materials and user information are managed with a standardized asset administration shell, offering advantages in later adjustment and versatility. This study presents an effective digital assistant system for industrial sites, innovating through large language models that interact dynamically with users to provide personalized supporting materials. A partial implementation of the materials for machine tools was carried out, including educational content for three levels using virtual reality, augmented reality, and generative natural language models. Additionally, the efficiency and effectiveness of the proposed system were evaluated through a comparison with widely used rule-based chatbots.

This paper aims to reduce potential human errors and loads that arise from the lack of novice human operators and different interfaces in advanced machine tool industries. A digital assistant using generative artificial intelligence is designed to answer questions about machine terminologies, and operation sequences happening in an alarm state. Combining Fine-Tuning, Retrieval Augmented Generation, and Prompt Engineering, it solves problems of common-purpose large language models. The proposed system is implemented on a local server and connected to a mobile device. It shows increasing quantitative accuracy from 51% to 79% and 84% in the fine-tuned and retriever model, and the qualitative score increases from 21 to 25 in the retriever model.

Food Safety Assurance Systems: Quality Assurance and Good Laboratory Practice

Food Safety Assurance Systems: Quality Assurance and Good Laboratory Practice