The reasoning capabilities of Large Language Models (LLMs) have increased greatly over the last few years, as have their size and complexity. Nonetheless, the use of LLMs in production remains challenging due to their unpredictable nature and discrepancies that can exist between their desired behavior and their actual model output. In this paper, we introduce a new tool, InTraVisTo (Inside Transformer Visualisation Tool), designed to enable researchers to investigate and trace the computational process that generates each token in a Transformer-based LLM. InTraVisTo provides a visualization of both the internal state of the Transformer model (by decoding token embeddings at each layer of the model) and the information flow between the various components across the different layers of the model (using a Sankey diagram). With InTraVisTo, we aim to help researchers and practitioners better understand the computations being performed within the Transformer model and thus to shed some light on internal patterns and reasoning processes employed by LLMs.

Effective storytelling relies on a delicate balance between meeting the reader's prior expectations and introducing unexpected developments. In the domain of detective fiction, this tension is known as fair play, which includes the implicit agreement between the writer and the reader as to the range of possible resolutions the mystery story may have. In this work, we present a probabilistic framework for detective fiction that allows us to define desired qualities. Using this framework, we formally define fair play and design appropriate metrics for it. Stemming from these definitions is an inherent tension between the coherence of the story, which measures how much it ``makes sense'', and the surprise it induces. We validate the framework by applying it to LLM-generated detective stories. This domain is appealing since we have an abundance of data, we can sample from the distribution generating the story, and the story-writing capabilities of LLMs are interesting in their own right. Results show that while LLM-generated stories may be unpredictable, they generally fail to balance the trade-off between surprise and fair play, which greatly contributes to their poor quality.

When deciding to read an article or incorporate it into their research, scholars often seek to quickly identify and understand its main ideas. In this paper, we aim to extract these key concepts and contributions from scientific articles in the form of Question and Answer (QA) pairs. We propose two distinct approaches for generating QAs. The first approach involves selecting salient paragraphs, using a Large Language Model (LLM) to generate questions, ranking these questions by the likelihood of obtaining meaningful answers, and subsequently generating answers. This method relies exclusively on the content of the articles. However, assessing an article's novelty typically requires comparison with the existing literature. Therefore, our second approach leverages a Knowledge Graph (KG) for QA generation. We construct a KG by fine-tuning an Entity Relationship (ER) extraction model on scientific articles and using it to build the graph. We then employ a salient triplet extraction method to select the most pertinent ERs per article, utilizing metrics such as the centrality of entities based on a triplet TF-IDF-like measure. This measure assesses the saliency of a triplet based on its importance within the article compared to its prevalence in the literature. For evaluation, we generate QAs using both approaches and have them assessed by Subject Matter Experts (SMEs) through a set of predefined metrics to evaluate the quality of both questions and answers. Our evaluations demonstrate that the KG-based approach effectively captures the main ideas discussed in the articles. Furthermore, our findings indicate that fine-tuning the ER extraction model on our scientific corpus is crucial for extracting high-quality triplets from such documents.

To overcome these significant challenges, we propose two novel architectures designed to integrate domain knowledge about drug side effects into a Llama 3 - 8B Language Model: Retrieval Augmented Generation (RAG) and GraphRAG. Our first architecture employs RAG, which enhances LLMs by retrieving relevant information from an external Pinecone vector database where drug side effect information is stored as feature vectors. The second architecture utilizes GraphRAG, which leverages a Neo4j graph database to stor e and efficiently handle more complex relationships of drug side effect associations. Both frameworks incorporate custom split functions and filtering modules to optimize user prompts for accurate retrieval. Through extensive evaluations on 19,520 associat ions between 976 marketed drugs and 3,851 unique side effect terms, we demonstrate that GraphRAG achieves near - perfect accuracy in drug side effect retrieval, significantly outperforming standalone LLMs and standard RAG approaches.

We propose a novel framework for open-ended video question answering that enhances reasoning depth and robustness in complex real-world scenarios, as benchmarked on the CVRR-ES dataset. Existing Video-Large Multimodal Models (Video-LMMs) often exhibit limited contextual understanding, weak temporal modeling, and poor generalization to ambiguous or compositional queries. To address these challenges, we introduce a prompting-and-response integration mechanism that coordinates multiple heterogeneous Video-Language Models (VLMs) via structured chains of thought, each tailored to distinct reasoning pathways. An external Large Language Model (LLM) serves as an evaluator and integrator, selecting and fusing the most reliable responses. Extensive experiments demonstrate that our method significantly outperforms existing baselines across all evaluation metrics, showcasing superior generalization and robustness. Our approach offers a lightweight, extensible strategy for advancing multimodal reasoning without requiring model retraining, setting a strong foundation for future Video-LMM development.

Large Language Models (LLMs) have demonstrated considerable potential in improving coding education by providing support for code writing, explanation, and debugging. However, existing LLM-based approaches generally fail to assess students' abilities, design learning plans, provide personalized material aligned with individual learning goals, and enable interactive learning. Current work mostly uses single LLM agents, which limits their ability to understand complex code repositories and schedule step-by-step tutoring. Recent research has shown that multi-agent LLMs can collaborate to solve complicated problems in various domains like software engineering, but their potential in the field of education remains unexplored. In this work, we introduce CodeEdu, an innovative multi-agent collaborative platform that combines LLMs with tool use to provide proactive and personalized education in coding. Unlike static pipelines, CodeEdu dynamically allocates agents and tasks to meet student needs. Various agents in CodeEdu undertake certain functions specifically, including task planning, personalized material generation, real-time QA, step-by-step tutoring, code execution, debugging, and learning report generation, facilitated with extensive external tools to improve task efficiency. Automated evaluations reveal that CodeEdu substantially enhances students' coding performance.

Large Language Models (LLMs) frequently reproduce the gender- and sexual-identity prejudices embedded in their training corpora, leading to outputs that marginalize LGBTQIA+ users. Hence, reducing such biases is of great importance. To achieve this, we evaluate two parameter-efficient fine-tuning (PEFT) techniques - Low-Rank Adaptation (LoRA) and soft-prompt tuning - as lightweight alternatives to full-model fine-tuning for mitigating such biases. Using the WinoQueer benchmark, we quantify bias in three open-source LLMs and observe baseline bias scores reaching up to 98 (out of 100) across a range of queer identities defined by gender and/or sexual orientation, where 50 would indicate neutrality. Fine-tuning with LoRA (< 0.1% additional parameters) on a curated QueerNews corpus reduces those scores by up to 50 points and raises neutrality from virtually 0% to as much as 36%. Soft-prompt tuning (10 virtual tokens) delivers only marginal improvements. These findings show that LoRA can deliver meaningful fairness gains with minimal computation. We advocate broader adoption of community-informed PEFT, the creation of larger queer-authored corpora, and richer evaluation suites beyond WinoQueer, coupled with ongoing audits to keep LLMs inclusive.

Few-shot class-incremental learning (FSCIL) is challenging due to extremely limited training data; while aiming to reduce catastrophic forgetting and learn new information. We propose Diffusion-FSCIL, a novel approach that employs a text-to-image diffusion model as a frozen backbone. Our conjecture is that FSCIL can be tackled using a large generative model's capabilities benefiting from 1) generation ability via large-scale pre-training; 2) multi-scale representation; 3) representational flexibility through the text encoder. To maximize the representation capability, we propose to extract multiple complementary diffusion features to play roles as latent replay with slight support from feature distillation for preventing generative biases. Our framework realizes efficiency through 1) using a frozen backbone; 2) minimal trainable components; 3) batch processing of multiple feature extractions. Extensive experiments on CUB-200, \emph{mini}ImageNet, and CIFAR-100 show that Diffusion-FSCIL surpasses state-of-the-art methods, preserving performance on previously learned classes and adapting effectively to new ones.

Rich and context-aware activity logs facilitate user behavior analysis and health monitoring, making them a key research focus in ubiquitous computing. The remarkable semantic understanding and generation capabilities of Large Language Models (LLMs) have recently created new opportunities for activity log generation. However, existing methods continue to exhibit notable limitations in terms of accuracy, efficiency, and semantic richness. To address these challenges, we propose DailyLLM. To the best of our knowledge, this is the first log generation and summarization system that comprehensively integrates contextual activity information across four dimensions: location, motion, environment, and physiology, using only sensors commonly available on smartphones and smartwatches. To achieve this, DailyLLM introduces a lightweight LLM-based framework that integrates structured prompting with efficient feature extraction to enable high-level activity understanding. Extensive experiments demonstrate that DailyLLM outperforms state-of-the-art (SOTA) log generation methods and can be efficiently deployed on personal computers and Raspberry Pi. Utilizing only a 1.5B-parameter LLM model, DailyLLM achieves a 17% improvement in log generation BERTScore precision compared to the 70B-parameter SOTA baseline, while delivering nearly 10x faster inference speed.

This study examines the role of human judges in legal decision-making by using machine learning to predict child physical custody outcomes in French appellate courts. Building on the legal realism-formalism debate, we test whether individual judges' decision-making patterns significantly influence case outcomes, challenging the assumption that judges are neutral variables that apply the law uniformly. To ensure compliance with French privacy laws, we implement a strict pseudonymization process. Our analysis uses 18,937 living arrangements rulings extracted from 10,306 cases. We compare models trained on individual judges' past rulings (specialist models) with a judge-agnostic model trained on aggregated data (generalist models). The prediction pipeline is a hybrid approach combining large language models (LLMs) for structured feature extraction and ML models for outcome prediction (RF, XGB and SVC). Our results show that specialist models consistently achieve higher predictive accuracy than the general model, with top-performing models reaching F1 scores as high as 92.85%, compared to the generalist model's 82.63% trained on 20x to 100x more samples. Specialist models capture stable individual patterns that are not transferable to other judges. In-Domain and Cross-Domain validity tests provide empirical support for legal realism, demonstrating that judicial identity plays a measurable role in legal outcomes. All data and code used will be made available.