Résultats de recherche pour « LLMs »

Large language models (LLMs) are increasingly used for automated news credibility assessment, yet it remains unclear whether they apply even-handed standards across journalistic genres. We examine whether zero-shot LLMs are more likely to misclassify legitimate entertainment news as fake than legitimate hard news, using a within-dataset design on GossipCop from FakeNewsNet. Across four frontier models, we find a clear but model-specific genre asymmetry: DeepSeek-V3.2 and GPT-5.2 show false-positive-rate gaps of 10.1 and 8.8 percentage points, respectively (both $p < .001$), whereas Claude Opus 4.6 and Gemini 3 Flash show no comparable difference. A style-swap experiment yields only limited and inconsistent changes, suggesting that the asymmetry is not reducible to stylistic register alone. Prompt-based mitigation is likewise possible but not generic: framing the model as an entertainment-news fact-checker reduces false positives for DeepSeek-V3.2 by about 50\% without detectable recall loss, but offers little improvement for GPT-5.2. Exploratory qualitative coding further suggests two recurring error patterns in sampled false positives: treating private-life claims as inherently unverifiable and discounting entertainment journalism as an epistemically weaker genre. Taken together, these findings show that aggregate performance metrics can obscure structured false positives within legitimate journalism. We argue that LLM-based credibility assessment may not only evaluate truth claims but also differentially recognize the legitimacy of journalistic genres, and that evaluation should therefore include genre-stratified false-positive analysis alongside overall accuracy.

We will introduce Large Language Models (LLMs), focusing on their core architecture, training methodologies, practical applications, and some important considerations. We begin with a brief exploration of the transformer architecture, examining how architectural choices and scaling laws have shaped modern language models. We then investigate key training paradigms, with particular emphasis on Reinforcement Learning from Human Feedback (RLHF) and instruction tuning, while briefly touching on emerging approaches like Direct Preference Optimisation (DPO). The discussion progresses through the evolution of knowledge integration techniques, from basic context stuffing to Retrieval-Augmented Generation (RAG), highlighting their practical implications. We will discuss the critical issue of hallucinations in LLM outputs, providing concrete examples and verification strategies. To conclude, a quick overview of the current software landscape and immediate future developments in the field, and a look ahead to the rest of the webinar series.

Large language models (LLMs) have become increasingly sophisticated, leading to widespread deployment in sensitive applications where safety and reliability are paramount. However, LLMs have inherent risks accompanying them, including bias, potential for unsafe actions, dataset poisoning, lack of explainability, hallucinations, and non-reproducibility. These risks necessitate the development of "guardrails" to align LLMs with desired behaviors and mitigate potential harm. This work explores the risks associated with deploying LLMs and evaluates current approaches to implementing guardrails and model alignment techniques. We examine intrinsic and extrinsic bias evaluation methods and discuss the importance of fairness metrics for responsible AI development. The safety and reliability of agentic LLMs (those capable of real-world actions) are explored, emphasizing the need for testability, fail-safes, and situational awareness. Technical strategies for securing LLMs are presented, including a layered protection model operating at external, secondary, and internal levels. System prompts, Retrieval-Augmented Generation (RAG) architectures, and techniques to minimize bias and protect privacy are highlighted. Effective guardrail design requires a deep understanding of the LLM's intended use case, relevant regulations, and ethical considerations. Striking a balance between competing requirements, such as accuracy and privacy, remains an ongoing challenge. This work underscores the importance of continuous research and development to ensure the safe and responsible use of LLMs in real-world applications.

Recent advances in large language models (LLMs) have led to increasingly strong claims abouttheir human-level linguistic competence and, by extension, their relevance for understanding naturallanguage and cognition. While a growing interdisciplinary literature has used behavioural tests andpsychometric tools to assess LLM performance, particularly in linguistics, such comparisons raisesubstantial conceptual and methodological concerns. This paper examines what current semanticlimitations in LLMs reveal about the nature of meaning in natural language. Drawing on empiricalfindings from psycholinguistics, the analysis shows that, despite fluent surface-level performance,LLMs systematically lack core aspects of semantic competence, including real-world grounding,communicative intent, and stable grammatical judgment. These limitations are then situated withinbroader theoretical frameworks from the philosophy of cognitive science, focusing on Miracchi’sreformulation of the Frame Problem and the Relevance Realisation framework. Together, theseapproaches highlight the central role of embodied agency, environmental embeddedness, and socio-cultural context in human meaning-making, and clarify why current computational architecturesstruggle with semantic relevance. The paper concludes by arguing for a bio-cultural conception oflanguage, suggesting that the semantic gaps observed in LLMs are not merely technicalshortcomings but rather reflect fundamental differences between algorithmic systems and human linguisticcognition.

Humans are believed to perceive numbers on a logarithmic mental number line, where smaller values are represented with greater resolution than larger ones. This cognitive bias, supported by neuroscience and behavioral studies, suggests that numerical magnitudes are processed in a sublinear fashion rather than on a uniform linear scale. Inspired by this hypothesis, we investigate whether large language models (LLMs) exhibit a similar logarithmic-like structure in their internal numerical representations. By analyzing how numerical values are encoded across different layers of LLMs, we apply dimensionality reduction techniques such as PCA and PLS followed by geometric regression to uncover latent structures in the learned embeddings. Our findings reveal that the model's numerical representations exhibit sublinear spacing, with distances between values aligning with a logarithmic scale. This suggests that LLMs, much like humans, may encode numbers in a compressed, non-uniform manner.

Phishing remains a persistent social engineering threat, and Large Language Models (LLMs) now play a dual role by enabling the generation of phishing-like messages while also being used as detectors. However, current evaluations often rely on static benchmark emails and do not isolate how controlled semantic variation affects LLM-based detection. We present a controlled framework that conditions LLM generation on semantic and lexical features extracted from real phishing emails, including topics, keywords, and synonyms, and evaluates multiple open-source LLMs as binary phishing detectors under standardized prompts and deterministic inference. The framework compares aligned and abliterated models for phishing generation and additionally examines their self-detection capabilities. Our results show that variants generated from a single conditioning feature produce higher false-negative rates than those generated from multiple features. Increasing the number of conditioning features significantly reduces the odds of false-negative classifications up to three features, while adding a fourth provides no significant further improvement. Self-detection is also unreliable across models. Abliteration has mixed, model-dependent effects on the evasiveness of generated phishing messages, but it generally reduces defensive reliability when the same models are used as detectors, increasing benign-email misclassification and, in some settings, substantially degrading discrimination. The proposed framework provides a reproducible methodology for systematically studying how controlled semantic variation influences the robustness of LLM-based phishing detection.

Prior research has demonstrated noticeable performance gains through the use of probabilistic tokenizations, an approach that involves employing multiple tokenizations of the same input string during the training phase of a language model. Despite these promising findings, modern large language models (LLMs) have yet to be trained using probabilistic tokenizations. Interestingly, while the tokenizers of these contemporary LLMs have the capability to generate multiple tokenizations, this property remains underutilized. In this work, we propose a novel method to leverage the multiple tokenization capabilities of modern LLM tokenizers, aiming to enhance the self-consistency of LLMs in reasoning tasks. Our experiments indicate that when utilizing probabilistic tokenizations, LLMs generate logically diverse reasoning paths, moving beyond mere surface-level linguistic diversity.We carefully study probabilistic tokenization and offer insights to explain the self consistency improvements it brings through extensive experimentation on 5 LLM families and 4 reasoning benchmarks.

Join us for a live panel discussion as part of our ongoing webinar series, Large Language Models and their applications in bioinformatics. This interactive QA session offers attendees the opportunity to engage with speakers from the series, ask new questions, and revisit key discussions.

Time series forecasting remains a challenging task, particularly in the context of complex multiscale temporal patterns. This study presents LLM-Mixer, a framework that improves forecasting accuracy through the combination of multiscale time-series decomposition with pre-trained LLMs (Large Language Models). LLM-Mixer captures both short-term fluctuations and long-term trends by decomposing the data into multiple temporal resolutions and processing them with a frozen LLM, guided by a textual prompt specifically designed for time-series data. Extensive experiments conducted on multivariate and univariate datasets demonstrate that LLM-Mixer achieves competitive performance, outperforming recent state-of-the-art models across various forecasting horizons. This work highlights the potential of combining multiscale analysis and LLMs for effective and scalable time-series forecasting.

This webinar will explore how Large Language Models (LLMs) can streamline the extraction of critical information from scientific texts, focusing on patient-derived cancer models (PDCMs). PDCMs are vital tools for cancer research and preclinical studies, with a growing body of literature in this field. However, manually extracting and curating information from scientific texts is labor-intensive and prone to delays. In this session, we will introduce two innovative approaches: direct prompting and soft prompting. Direct prompting uses manually created instructions to extract PDCM-related entities, while soft prompting leverages machine learning to train continuous vector prompts. We will discuss our comparative evaluation using state-of-the-art proprietary and open-source LLMs, demonstrating how tailored prompt engineering can elevate the performance of smaller, open models to match proprietary counterparts. This session will highlight the potential of LLMs to enhance domain-specific knowledge extraction and accelerate research workflows.