Benchmarking Knowledge Boundary for Large Language Models: A Different Perspective on Model Evaluation

Biases in language and their reflection in language models have attracted researchers' attention, particularly with the growth of large language models (LLMs). However, many questions on the links between language models and people’s biased attitudes remain unanswered. In the current study we focus on gender–career bias to examine the extent to which language models can be used to model behavioural responses in the Gender–Career Implicit Association Test (IAT). We provide a systematic evaluation of a range of language models, including n-gram, count vector, predict (word2vec), and Large Language Models (LLMs), to determine how well they capture people’s behaviour in the IAT. We compared response time data from over 800,000 participants against 25 language models, with a total of 675 model variants. We find that many language models, including large language models (LLMs), correlated well with human behavior. While results support previous findings for both predict and count model families, we observed that performance of LLMs was consistently different from that of simpler predict models, particularly in terms of the direction and strength of correlations with reaction time and bias. This divergence may indicate successful attempts to mitigate bias in LLMs while preserving other aspects of linguistic information. Our findings reinforce the idea that societal biases are generally encoded in language, but that large language models can exhibit behaviors different to classical language models.

Purpose This study aims to explore the integration of large language models (LLMs) and vision-language models (VLMs) in robotics, highlighting their potential benefits and the safety challenges they introduce, including robustness issues, adversarial vulnerabilities, privacy concerns and ethical implications. Design/methodology/approach This survey conducts a comprehensive analysis of the safety risks associated with LLM- and VLM-powered robotic systems. The authors review existing literature, analyze key challenges, evaluate current mitigation strategies and propose future research directions. Findings The study identifies that ensuring the safety of LLM-/VLM-driven robots requires a multi-faceted approach. While current mitigation strategies address certain risks, gaps remain in real-time monitoring, adversarial robustness and ethical safeguards. Originality/value This study offers a structured and comprehensive overview of the safety challenges in LLM-/VLM-driven robotics. It contributes to ongoing discussions by integrating technical, ethical and regulatory perspectives to guide future advancements in safe and responsible artificial intelligence-driven robotics.

The subject of the study is the analysis and improvement of methods for determining the relevance of project names to the information content of purchases using large language models. The object of the study is a database containing the names of projects and purchases in the field of electric power industry, collected from open sources. The author examines in detail such aspects of the topic as the use of TF-IDF and cosine similarity metrics for primary data filtering, and also describes in detail the integration and evaluation of the effectiveness of large language models such as GigaChat, GPT-3.5, and GPT-4 in text data matching tasks. Special attention is paid to the methods of clarifying the similarity of names based on reflection introduced into the prompta of large language models, which makes it possible to increase the accuracy of data comparison. The study uses TF-IDF and cosine similarity methods for primary data analysis, as well as large GigaChat, GPT-3.5 and GPT-4 language models for detailed verification of the relevance of project names and purchases, including reflection in model prompta to improve the accuracy of results. The novelty of the research lies in the development of a combined approach to determining the relevance of project names and purchases, combining traditional methods of processing text information (TF-IDF, cosine similarity) with the capabilities of large language models. A special contribution of the author to the research of the topic is the proposed methodology for improving the accuracy of data comparison by clarifying the results of primary selection using GPT-3.5 and GPT-4 models with optimized prompta, including reflection. The main conclusions of the study are confirmation of the prospects of using the developed approach in the tasks of information support for procurement processes and project implementation, as well as the possibility of using the results obtained for the development of text data mining systems in various sectors of the economy. The study showed that the use of language models makes it possible to improve the value of the F2 measure to 0.65, which indicates a significant improvement in the quality of data comparison compared with basic methods.

The application of deep learning to tabular data remains an ongoing challenge, with tree-based models such as XGBoost consistently outperforming neural network methods in most real-world scenarios. Recent advances in Large Language Models (LLMs) highlight their ability to generalize to new tasks via few-shot or one-shot prompting, yet questions remain about their utility in structured, tabular domains—particularly for resource-efficient Small Language Models (SLMs). In this work, we systematically evaluate the effectiveness of both large and small transformer-based language models for tabular classification tasks using few-shot strategies. Our approach investigates input serialization schemes and prompt engineering to maximize performance in low-data regimes. Experiments on benchmark datasets—including Diabetes, Heart Failure, and German Credit Risk—demonstrate that well-designed SLMs can approach, and occasionally match, the performance of much larger models, substantially reducing the need for extensive labeled data and retraining. However, further advances are required for language models to consistently rival the strongest tree-ensemble baselines. Our findings support the idea that SLMs, when properly prompted, offer a promising, flexible, and label-efficient alternative for automating and dynamizing machine learning pipelines on tabular data.

The paper studies the problem of software code migration during the transition from the old version of the SAP ERP 6.0 system to modern ERP platforms, in particular SAP S/4HANA. The necessity of automation of the migration process using artificial intelligence methods and tools is substantiated. A comprehensive analysis of existing academic and industrial case studies confirmed the feasibility of using large language models for code migration. A number of advantages of the approach based on large language models are revealed, in particular, the ability to handle various migration scenarios. The key risks and limitations common in the enterprise environment are highlighted, including data security, privacy, high cost of API calls and limited knowledge of models regarding the subject area. A method is developed for selecting a large open source language model that can be safely used in the enterprise landscape, thereby mitigating the risks associated with cloud solutions. The method is based on the developed system of criteria, including: model size, context window size, license type, support for fine-tuning training, focus on code quality (measured using a reference dataset), support for model quantization and community support. The hybrid AHP-TOPSIS method is proposed for systematic ranking of candidate models. The AHP method was used to check the weights of the criteria, while TOPSIS was used to rank the models based on their proximity to the ideal solution. Based on the developed method, three large language models were selected: Qwen 2.5 Coder 14B, DeepSeek-Coder-V2 16B and Llama 3.1 8B. A comprehensive testing plan for the selected models was developed for further experimental studies. The scientific novelty of the work lies in the development of a method for selecting large language models for custom code migration tasks, which differs from the existing system of model selection criteria and the use of a hybrid AHP-TOPSIS approach for ranking candidate models. The practical value of the work is that the developed method allows you to select a model taking into account the constraints of the corporate environment and information security requirements, which will increase the level of automation in the migration of AVAP when switching to new versions of ERP systems from SAP SE. The developed method can be applied to the selection of large language models when upgrading any large computer systems developed in both open source and proprietary programming languages.

Large Language Models (LLMs) have enormous potential in healthcare, but we also need to be careful and implement strong safety measures. These advanced computer systems can understand complex medical questions written in everyday language and often provide accurate medical information. This article explores how the field has evolved, from the first language models to the LLMs we have now, specifically designed or adjusted for use in the medical world. We explore how LLMs can make important language-related tasks in healthcare more efficient and effective, such as identifying medical terms, understanding how different pieces of information are connected, drawing conclusions from written text, sorting documents, and answering medical queries. We also shed light on new and exciting ways LLMs can use more than just text, like combining written information with images or other types of data. We compare some of the best recent LLMs and discuss how they can be used in everyday medical situations. We also measure how well these models perform in the biomedical field, explaining what each measurement tells us and where it falls short. Lastly, by discussing the biggest hurdles LLMs face in healthcare—things like generating incorrect information (hallucinations), unfair biases, protecting patient privacy, making sure the models are dependable, and fitting these new tools into how doctors and nurses already work. We also provide a roadmap for future research. We aim to present a clear and fair picture of what LLMs can offer healthcare right now and what's necessary to use them safely and dependably. This includes explaining how researchers are working to make these models easier for doctors and patients to understand. We followed the PRISMA research method to guide our review of the field, from early language models to current LLMs adapted for clinical use.

Large language models (LLMs) have been proven capable of memorizing their training data, which can be extracted through specifically designed prompts. As the scale of datasets continues to grow, privacy risks arising from memorization have attracted increasing attention. Quantifying language model memorization helps evaluate potential privacy risks. However, prior works on quantifying memorization require access to the precise original data or incur substantial computational overhead, making it difficult for applications in real-world language models. To this end, we propose a fine-grained, entity-level definition to quantify memorization with conditions and metrics closer to real-world scenarios. In addition, we also present an approach for efficiently extracting sensitive entities from autoregressive language models. We conduct extensive experiments based on the proposed, probing language models' ability to reconstruct sensitive entities under different settings. We find that language models have strong memorization at the entity level and are able to reproduce the training data even with partial leakages. The results demonstrate that LLMs not only memorize their training data but also understand associations between entities. These findings necessitate that trainers of LLMs exercise greater prudence regarding model memorization, adopting memorization mitigation techniques to preclude privacy violations.

There is much excitement about the opportunity to harness the power of large language models (LLMs) when building problem-solving assistants. However, the standard methodology of evaluating LLMs relies on static pairs of inputs and outputs; this is insufficient for making an informed decision about which LLMs are best to use in an interactive setting, and how that varies by setting. Static assessment therefore limits how we understand language model capabilities. We introduce CheckMate, an adaptable prototype platform for humans to interact with and evaluate LLMs. We conduct a study with CheckMate to evaluate three language models (InstructGPT, ChatGPT, and GPT-4) as assistants in proving undergraduate-level mathematics, with a mixed cohort of participants from undergraduate students to professors of mathematics. We release the resulting interaction and rating dataset, MathConverse. By analyzing MathConverse, we derive a taxonomy of human query behaviors and uncover that despite a generally positive correlation, there are notable instances of divergence between correctness and perceived helpfulness in LLM generations, among other findings. Further, we garner a more granular understanding of GPT-4 mathematical problem-solving through a series of case studies, contributed by experienced mathematicians. We conclude with actionable takeaways for ML practitioners and mathematicians: models that communicate uncertainty, respond well to user corrections, and can provide a concise rationale for their recommendations, may constitute better assistants. Humans should inspect LLM output carefully given their current shortcomings and potential for surprising fallibility.

When fact-checking methods based on large language models (LLMs) use external evidence to validate claims, knowledge conflicts often arise. These conflicts typically stem from inconsistencies between the external evidence and LLMs' internal pre-existing knowledge. Such an inconsistency could lead LLMs to draw incorrect answers when validating claims, especially when they are overly confident in their internal incorrect knowledge. Previous works on LLM-based fact-checking have overlooked this issue. This paper, for the first time, proposes a framework (namely KnowFC) to navigate this issue. Our key insight is dividing and adaptively utilizing the knowledge that LLMs know and do not know, thereby avoiding conflicts while enhancing the correctness and efficiency of fact-checking. Specifically, in KnowFC, we propose an adaptive retrieval method, where we train an LLM using a reinforcement learning algorithm coupled with the Dunning-Kruger effect-inspired reward mechanism to identify its knowledge boundaries through confidence calibration, thereby realizing adaptive evidence retrieval. Besides, we propose a reliable and debiased fact verification method, where we organize and construct reasoning graphs using retrieved evidence to verify claims, followed by a causal intervention method using causal mediation analysis to mitigate internal knowledge interference. Experimental results on both FEVEROUS and AVeriTeC datasets show that our method outperforms baseline methods in terms of accuracy and F1 score, while also improving fact-checking efficiency.

Several recent works seek to develop foundation models specifically for medical applications, adapting general-purpose large language models (LLMs) and vision-language models (VLMs) via continued pretraining on publicly available biomedical corpora.These works typically claim that such domain-adaptive pretraining (DAPT) improves performance on downstream medical tasks, such as answering medical licensing exam questions.In this paper, we compare seven public "medical" LLMs and two VLMs against their corresponding base models, arriving at a different conclusion: all medical VLMs and nearly all medical LLMs fail to consistently improve over their base models in the zero-/few-shot prompting regime for medical question-answering (QA) tasks.For instance, across the tasks and model pairs we consider in the 3-shot setting, medical LLMs only outperform their base models in 12.1% of cases, reach a (statistical) tie in 49.8% of cases, and are significantly worse than their base models in the remaining 38.2% of cases.Our conclusions are based on (i) comparing each medical model head-to-head, directly against the corresponding base model; (ii) optimizing the prompts for each model separately; and (iii) accounting for statistical uncertainty in comparisons.While these basic practices are not consistently adopted in the literature, our ablations show that they substantially impact conclusions.Our findings suggest that state-of-the-art generaldomain models may already exhibit strong medical knowledge and reasoning capabilities, and offer recommendations to strengthen the conclusions of future studies.

Large language models (LLMs) have made a significant impact on the fields of general artificial intelligence. General purpose LLMs exhibit strong logic and reasoning skills and general world knowledge but can sometimes generate misleading results when prompted on specific subject areas. LLMs trained with domain-specific knowledge can reduce the generation of misleading information (i.e. hallucinations) and enhance the precision of LLMs in specialized contexts. Training new LLMs on specific corpora however can be resource intensive. Here we explored the use of a retrieval-augmented generation (RAG) model which we tested on literature specific to a biomedical research area. OpenAI's GPT-3.5, GPT-4, Microsoft's Prometheus, and a custom RAG model were used to answer 19 questions pertaining to diffuse large B-cell lymphoma (DLBCL) disease biology and treatment. Eight independent reviewers assessed LLM responses based on accuracy, relevance, and readability, rating responses on a 3-point scale for each category. These scores were then used to compare LLM performance. The performance of the LLMs varied across scoring categories. On accuracy and relevance, the RAG model outperformed other models with higher scores on average and the most top scores across questions. GPT-4 was more comparable to the RAG model on relevance versus accuracy. By the same measures, GPT-4 and GPT-3.5 had the highest scores for readability of answers when compared to the other LLMs. GPT-4 and 3.5 also had more answers with hallucinations than the other LLMs, due to non-existent references and inaccurate responses to clinical questions. Our findings suggest that an oncology research-focused RAG model may outperform general-purpose LLMs in accuracy and relevance when answering subject-related questions. This framework can be tailored to Q&A in other subject areas. Further research will help understand the impact of LLM architectures, RAG methodologies, and prompting techniques in answering questions across different subject areas.

In this article, we discuss ethical issues related to using and disclosing artificial intelligence (AI) tools, such as ChatGPT and other systems based on large language models (LLMs), to write or edit scholarly manuscripts. Some journals, such as Science , have banned the use of LLMs because of the ethical problems they raise concerning responsible authorship. We argue that this is not a reasonable response to the moral conundrums created by the use of LLMs because bans are unenforceable and would encourage undisclosed use of LLMs. Furthermore, LLMs can be useful in writing, reviewing and editing text, and promote equity in science. Others have argued that LLMs should be mentioned in the acknowledgments since they do not meet all the authorship criteria. We argue that naming LLMs as authors or mentioning them in the acknowledgments are both inappropriate forms of recognition because LLMs do not have free will and therefore cannot be held morally or legally responsible for what they do. Tools in general, and software in particular, are usually cited in-text, followed by being mentioned in the references. We provide suggestions to improve APA Style for referencing ChatGPT to specifically indicate the contributor who used LLMs (because interactions are stored on personal user accounts), the used version and model (because the same version could use different language models and generate dissimilar responses, e.g., ChatGPT May 12 Version GPT3.5 or GPT4), and the time of usage (because LLMs evolve fast and generate dissimilar responses over time). We recommend that researchers who use LLMs: (1) disclose their use in the introduction or methods section to transparently describe details such as used prompts and note which parts of the text are affected, (2) use in-text citations and references (to recognize their used applications and improve findability and indexing), and (3) record and submit their relevant interactions with LLMs as supplementary material or appendices.

Large language models have recently achieved state of the art performance across a wide variety of natural language tasks. Meanwhile, the size of these models and their latency have significantly increased, which makes their usage costly, and raises an interesting question: do language models need to be large? We study this question through the lens of model compression. We present a generic, structured pruning approach by parameterizing each weight matrix using its low-rank factorization, and adaptively removing rank-1 components during training. On language modeling tasks, our structured approach outperforms other unstructured and block-structured pruning baselines at various compression levels, while achieving significant speedups during both training and inference. We also demonstrate that our method can be applied to pruning adaptive word embeddings in large language models, and to pruning the BERT model on several downstream fine-tuning classification benchmarks.

With large-scale language models demonstrating superior capabilities in a wide range of downstream natural language processing tasks, the future trajectory of research in the field of text categorization faces increasing uncertainty. In this evolving paradigm of open-ended language modeling, where task delimitations are increasingly blurred, a pressing question arises: to what extent has text classification advanced under the full potential of large language model (LLM)? To address this pivotal inquiry, we introduce recurrent generative pre-trained transformer (RGPT), an adaptive boosting framework meticulously designed to craft a dedicated LLM for text classification. RGPT constructs a sequence of base learners by dynamically modulating the training data distribution and iteratively fine-tuning LLMs. These base learners are then progressively integrated, leveraging historical prediction trajectories to form a highly specialized text classification model. Extensive empirical evaluations demonstrate that RGPT surpasses eight state-of-the-art pretrained language models and seven cutting-edge LLMs across four benchmark datasets, achieving an average performance gain of 2.90%.

PurposeLarge language models (LLMs) like Generative Pre-trained Transformer 4 (GPT-4) can assist in detecting errors in radiology reports, but privacy concerns limit their clinical applicability. This study compares closed-source and privacy-compliant open-source LLMs for detecting common errors in radiology reports.Materials and methodsA total of 120 radiology reports were compiled (30 each from X-ray, ultrasound, CT, and MRI). Subsequently, 397 errors from five categories (typographical, numerical, findings-impression discrepancies, omission/insertion, interpretation) were inserted into 100 of these reports; 20 reports were left unchanged. Two open-source models (Llama 3-70b, Mixtral 8x22b) and two commercial closed-source (GPT-4, GPT-4o) were tasked with error detection using identical prompts. The Kruskall–Wallis test and paired t-test were used for statistical analysis.ResultsOpen-source LLMs required less processing time per radiology report than closed-source LLMs (6 ± 2 s vs. 13 ± 4 s; p < 0.001). Closed-source LLMs achieved higher error detection rates than open-source LLMs (GPT-4o: 88% [348/397; 95% CI: 86, 92], GPT-4: 83% [328/397; 95% CI: 80, 87], Llama 3-70b: 79% [311/397; 95% CI: 76, 83], Mixtral 8x22b: 73% [288/397; 95% CI: 68, 77]; p < 0.001). Numerical errors (88% [67/76; 95% CI: 82, 93]) were detected significantly more often than typographical errors (75% [65/86; 95% CI: 68, 82]; p = 0.02), discrepancies between findings and impression (73% [73/101; 95% CI: 67, 80]; p < 0.01), and interpretation errors (70% [50/71; 95% CI: 62, 78]; p = 0.001).ConclusionOpen-source LLMs demonstrated effective error detection, albeit with comparatively lower accuracy than commercial closed-source models, and have potential for clinical applications when deployed via privacy-compliant local hosting solutions.Key PointsQuestionCan privacy-compliant open-source large language models (LLMs) match the error-detection performance of commercial non-privacy-compliant closed-source models in radiology reports?FindingsClosed-source LLMs achieved slightly higher accuracy in detecting radiology report errors than open-source models, with Llama 3-70b yielding the best results among the open-source models.Clinical relevanceOpen-source LLMs offer a privacy-compliant alternative for automated error detection in radiology reports, improving clinical workflow efficiency while ensuring patient data confidentiality. Further refinement could enhance their accuracy, contributing to better diagnosis and patient care.Graphical