This technical report analyzes the challenge of "hallucinations" (false information) in LLMs applied to law. It examines their causes, manifestations, and the effectiveness of the RAG mitigation strategy, highlighting its limitations and proposing holistic optimizations. The paper explores the ethical and regulatory implications, emphasizing human oversight as an irreplaceable role. It concludes that the solution lies not in incrementally improving generative models, but in adopting a "consultative" AI paradigm that prioritizes veracity and traceability, acting as a tool to amplify, not replace, professional judgment. -- Este informe técnico analiza el desafío de las "alucinaciones" (información falsa) en los LLMs aplicados al derecho. Se examinan sus causas, manifestaciones y la efectividad de la estrategia de mitigación RAG, exponiendo sus limitaciones y proponiendo optimizaciones holísticas. Se exploran las implicaciones éticas y regulatorias, enfatizando la supervisión humana como un rol insustituible. El documento concluye que la solución no reside en mejorar incrementalmente los modelos generativos, sino en adoptar un paradigma de IA "consultiva" que priorice la veracidad y la trazabilidad, actuando como una herramienta para amplificar, y no sustituir, el juicio profesional.

Recent advances in image-level self-supervised learning (SSL) have made significant progress, yet learning dense representations for patches remains challenging. Mainstream methods encounter an over-dispersion phenomenon that patches from the same instance/category scatter, harming downstream performance on dense tasks. This work reveals that image-level SSL avoids over-dispersion by involving implicit semantic concentration. Specifically, the non-strict spatial alignment ensures intra-instance consistency, while shared patterns, i.e., similar parts of within-class instances in the input space, ensure inter-image consistency. Unfortunately, these approaches are infeasible for dense SSL due to their spatial sensitivity and complicated scene-centric data. These observations motivate us to explore explicit semantic concentration for dense SSL. First, to break the strict spatial alignment, we propose to distill the patch correspondences. Facing noisy and imbalanced pseudo labels, we propose a noise-tolerant ranking loss. The core idea is extending the Average Precision (AP) loss to continuous targets, such that its decision-agnostic and adaptive focusing properties prevent the student model from being misled. Second, to discriminate the shared patterns from complicated scenes, we propose the object-aware filter to map the output space to an object-based space. Specifically, patches are represented by learnable prototypes of objects via cross-attention. Last but not least, empirical studies across various tasks soundly support the effectiveness of our method. Code is available in https://github.com/KID-7391/CoTAP.

Reader curiosity, the drive to seek information, is crucial for textual engagement, yet remains relatively underexplored in NLP. Building on Loewenstein's Information Gap Theory, we introduce a framework that models reader curiosity by quantifying semantic information gaps within a text's semantic structure. Our approach leverages BERTopic-inspired topic modeling and persistent homology to analyze the evolving topology (connected components, cycles, voids) of a dynamic semantic network derived from text segments, treating these features as proxies for information gaps. To empirically evaluate this pipeline, we collect reader curiosity ratings from participants (n = 49) as they read S. Collins's ''The Hunger Games'' novel. We then use the topological features from our pipeline as independent variables to predict these ratings, and experimentally show that they significantly improve curiosity prediction compared to a baseline model (73% vs. 30% explained deviance), validating our approach. This pipeline offers a new computational method for analyzing text structure and its relation to reader engagement.

A recent line of research on spoken language assessment (SLA) employs neural models such as BERT and wav2vec 2.0 (W2V) to evaluate speaking proficiency across linguistic and acoustic modalities. Although both models effectively capture features relevant to oral competence, each exhibits modality-specific limitations. BERT-based methods rely on ASR transcripts, which often fail to capture prosodic and phonetic cues for SLA. In contrast, W2V-based methods excel at modeling acoustic features but lack semantic interpretability. To overcome these limitations, we propose a system that integrates W2V with Phi-4 multimodal large language model (MLLM) through a score fusion strategy. The proposed system achieves a root mean square error (RMSE) of 0.375 on the official test set of the Speak & Improve Challenge 2025, securing second place in the competition. For comparison, the RMSEs of the top-ranked, third-ranked, and official baseline systems are 0.364, 0.384, and 0.444, respectively.

To collaborate effectively with humans, language models must be able to explain their decisions in natural language. We study a specific type of self-explanation: self-generated counterfactual explanations (SCEs), where a model explains its prediction by modifying the input such that it would have predicted a different outcome. We evaluate whether LLMs can produce SCEs that are valid, achieving the intended outcome, and minimal, modifying the input no more than necessary. When asked to generate counterfactuals, we find that LLMs typically produce SCEs that are valid, but far from minimal, offering little insight into their decision-making behaviour. Worryingly, when asked to generate minimal counterfactuals, LLMs typically make excessively small edits that fail to change predictions. The observed validity-minimality trade-off is consistent across several LLMs, datasets, and evaluation settings. Our findings suggest that SCEs are, at best, an ineffective explainability tool and, at worst, can provide misleading insights into model behaviour. Proposals to deploy LLMs in high-stakes settings must consider the impact of unreliable self-explanations on downstream decision-making. Our code is available at https://github.com/HarryMayne/SCEs.

Recent advances in large language models (LLMs) have enabled the emergence of general-purpose agents for automating end-to-end machine learning (ML) workflows, including data analysis, feature engineering, model training, and competition solving. However, existing benchmarks remain limited in task coverage, domain diversity, difficulty modeling, and evaluation rigor, failing to capture the full capabilities of such agents in realistic settings. We present TAM Bench, a diverse, realistic, and structured benchmark for evaluating LLM-based agents on end-to-end ML tasks. TAM Bench features three key innovations: (1) A browser automation and LLM-based task acquisition system that automatically collects and structures ML challenges from platforms such as Kaggle, AIcrowd, and Biendata, spanning multiple task types and data modalities (e.g., tabular, text, image, graph, audio); (2) A leaderboard-driven difficulty modeling mechanism that estimates task complexity using participant counts and score dispersion, enabling scalable and objective task calibration; (3) A multi-dimensional evaluation framework incorporating performance, format compliance, constraint adherence, and task generalization. Based on 150 curated AutoML tasks, we construct three benchmark subsets of different sizes -- Lite, Medium, and Full -- designed for varying evaluation scenarios. The Lite version, with 18 tasks and balanced coverage across modalities and difficulty levels, serves as a practical testbed for daily benchmarking and comparative studies.

Classifying patents by their relevance to the UN Sustainable Development Goals (SDGs) is crucial for tracking how innovation addresses global challenges. However, the absence of a large, labeled dataset limits the use of supervised learning. Existing methods, such as keyword searches, transfer learning, and citation-based heuristics, lack scalability and generalizability. This paper frames patent-to-SDG classification as a weak supervision problem, using citations from patents to SDG-tagged scientific publications (NPL citations) as a noisy initial signal. To address its sparsity and noise, we develop a composite labeling function (LF) that uses large language models (LLMs) to extract structured concepts, namely functions, solutions, and applications, from patents and SDG papers based on a patent ontology. Cross-domain similarity scores are computed and combined using a rank-based retrieval approach. The LF is calibrated via a custom positive-only loss that aligns with known NPL-SDG links without penalizing discovery of new SDG associations. The result is a silver-standard, soft multi-label dataset mapping patents to SDGs, enabling the training of effective multi-label regression models. We validate our approach through two complementary strategies: (1) internal validation against held-out NPL-based labels, where our method outperforms several baselines including transformer-based models, and zero-shot LLM; and (2) external validation using network modularity in patent citation, co-inventor, and co-applicant graphs, where our labels reveal greater thematic, cognitive, and organizational coherence than traditional technological classifications. These results show that weak supervision and semantic alignment can enhance SDG classification at scale.

Knowledge graphs, a powerful tool for structuring information through relational triplets, have recently become the new front-runner in enhancing question-answering systems. While traditional Retrieval Augmented Generation (RAG) approaches are proficient in fact-based and local context-based extraction from concise texts, they encounter limitations when addressing the thematic and holistic understanding of complex, extensive texts, requiring a deeper analysis of both text and context. This paper presents a comprehensive technical comparative study of three different methodologies for constructing knowledge graph triplets and integrating them with Large Language Models (LLMs) for question answering: spaCy, Stanford CoreNLP-OpenIE, and GraphRAG, all leveraging open source technologies. We evaluate the effectiveness, feasibility, and adaptability of these methods by analyzing their capabilities, state of development, and their impact on the performance of LLM-based question answering. Experimental results indicate that while OpenIE provides the most comprehensive coverage of triplets, GraphRAG demonstrates superior reasoning abilities among the three. We conclude with a discussion on the strengths and limitations of each method and provide insights into future directions for improving knowledge graph-based question answering.

In this technical report, we describe our submission for Task 1, Low-Complexity Device-Robust Acoustic Scene Classification, of the DCASE 2025 Challenge. Our work tackles the dual challenges of strict complexity constraints and robust generalization to both seen and unseen devices, while also leveraging the new rule allowing the use of device labels at test time. Our proposed system is based on a knowledge distillation framework where an efficient CP-MobileNet student learns from a compact, specialized two-teacher ensemble. This ensemble combines a baseline PaSST teacher, trained with standard cross-entropy, and a 'generalization expert' teacher. This expert is trained using our novel Device-Aware Feature Alignment (DAFA) loss, adapted from prior work, which explicitly structures the feature space for device robustness. To capitalize on the availability of test-time device labels, the distilled student model then undergoes a final device-specific fine-tuning stage. Our proposed system achieves a final accuracy of 57.93\% on the development set, demonstrating a significant improvement over the official baseline, particularly on unseen devices.

Providing students with flexible and timely academic support is a challenge at most colleges and universities, leaving many students without help outside scheduled hours. Large language models (LLMs) are promising for bridging this gap, but interactions between students and LLMs are rarely overseen by educators. We developed and studied an LLM-powered course assistant deployed across multiple computer science courses to characterize real-world use and understand pedagogical implications. By Spring 2024, our system had been deployed to approximately 2,000 students across six courses at three institutions. Analysis of the interaction data shows that usage remains strong in the evenings and nights and is higher in introductory courses, indicating that our system helps address temporal support gaps and novice learner needs. We sampled 200 conversations per course for manual annotation: most sampled responses were judged correct and helpful, with a small share unhelpful or erroneous; few responses included dedicated examples. We also examined an inquiry-based learning strategy: only around 11% of sampled conversations contained LLM-generated follow-up questions, which were often ignored by students in advanced courses. A Bloom's taxonomy analysis reveals that current LLM capabilities are limited in generating higher-order cognitive questions. These patterns suggest opportunities for pedagogically oriented LLM-based educational systems and greater educator involvement in configuring prompts, content, and policies.