Model compression offers a promising path to reducing the cost and inaccessibility of large pre-trained models, without significantly compromising their impressive performance. Large Transformer models, including large language models (LLMs), often contain computational redundancy, which can serve as a target for new model compression methods. In this work, we specifically target neuron-level redundancies in model layers by combining groups of similar neurons into fewer neurons. We frame this width reduction as a Discrete Optimal Transport problem, and propose DOTResize, a novel Transformer compression method that uses optimal transport theory to transform and compress model weights. To ensure applicability within the Transformer architecture, we motivate and incorporate entropic regularization and matrix factorization into the transportation maps produced by our method. Unlike pruning-based approaches which discard neurons based on importance measures, DOTResize re-projects the entire neuron width, allowing the retention and redistribution of useful signal across the reduced layer. Empirical results show that compared to simple or state-of-the-art neuron width-pruning techniques, DOTResize can outperform these methods across multiple LLM families and sizes, while achieving measurable reductions in real-world computational cost.

Large language models (LLMs) are rapidly being integrated into psychological research as research tools, evaluation targets, human simulators, and cognitive models. However, recent evidence reveals severe measurement unreliability: Personality assessments collapse under factor analysis, moral preferences reverse with punctuation changes, and theory-of-mind accuracy varies widely with trivial rephrasing. These "measurement phantoms"--statistical artifacts masquerading as psychological phenomena--threaten the validity of a growing body of research. Guided by the dual-validity framework that integrates psychometrics with causal inference, we present a six-stage workflow that scales validity requirements to research ambition--using LLMs to code text requires basic reliability and accuracy, while claims about psychological properties demand comprehensive construct validation. Researchers must (1) explicitly define their research goal and corresponding validity requirements, (2) develop and validate computational instruments through psychometric testing, (3) design experiments that control for computational confounds, (4) execute protocols with transparency, (5) analyze data using methods appropriate for non-independent observations, and (6) report findings within demonstrated boundaries and use results to refine theory. We illustrate the workflow through an example of model evaluation--"LLM selfhood"--showing how systematic validation can distinguish genuine computational phenomena from measurement artifacts. By establishing validated computational instruments and transparent practices, this workflow provides a path toward building a robust empirical foundation for AI psychology research.

While attribution methods, such as Shapley values, are widely used to explain the importance of features or training data in traditional machine learning, their application to Large Language Models (LLMs), particularly within Retrieval-Augmented Generation (RAG) systems, is nascent and challenging. The primary obstacle is the substantial computational cost, where each utility function evaluation involves an expensive LLM call, resulting in direct monetary and time expenses. This paper investigates the feasibility and effectiveness of adapting Shapley-based attribution to identify influential retrieved documents in RAG. We compare Shapley with more computationally tractable approximations and some existing attribution methods for LLM. Our work aims to: (1) systematically apply established attribution principles to the RAG document-level setting; (2) quantify how well SHAP approximations can mirror exact attributions while minimizing costly LLM interactions; and (3) evaluate their practical explainability in identifying critical documents, especially under complex inter-document relationships such as redundancy, complementarity, and synergy. This study seeks to bridge the gap between powerful attribution techniques and the practical constraints of LLM-based RAG systems, offering insights into achieving reliable and affordable RAG explainability.

--Large language models (LLMs) have transformed natural language processing, but their ability to memorize training data poses significant privacy risks. This paper investigates model inversion attacks on the Llama 3.2 model, a multilingual LLM developed by Meta. By querying the model with carefully crafted prompts, we demonstrate the extraction of personally identifiable information (PII) such as passwords, email addresses, and account numbers. Our findings highlight the vulnerability of even smaller LLMs to privacy attacks and underscore the need for robust defenses. We discuss potential mitigation strategies, including differential privacy and data sanitization, and call for further research into privacy-preserving machine learning techniques.

The rapid growth of large language models (LLMs) with traditional centralized fine-tuning emerges as a key technique for adapting these models to domain-specific challenges, yielding privacy risks for both model and data owners. One promising solution, called offsite-tuning (OT), is proposed to address these challenges, where a weaker emulator is compressed from the original model and further fine-tuned with adapter to enhance privacy. However, the existing OT-based methods require high computational costs and lack theoretical analysis. This paper introduces a novel OT approach based on gradient-preserving compression, named GradOT. By analyzing the OT problem through the lens of optimization, we propose a method that selectively applies compression techniques such as rank compression and channel pruning, preserving the gradients of fine-tuned adapters while ensuring privacy. Extensive experiments demonstrate that our approach surpasses existing OT methods, both in terms of privacy protection and model performance. Our method provides a theoretical foundation for OT and offers a practical, training-free solution for offsite-tuning of large-scale LLMs.

The onset of spontaneous thoughts are reflective of dynamic interactions between cognition, emotion, and attention. Typically, these experiences are studied through subjective appraisals that focus on their triggers, phenomenology, and emotional salience. In this work, we use linguistic signatures to investigate D ej ` a Vu, Involuntary Autobiographical Memories, and Unexpected Thoughts. Specifically, we analyze the inherent characteristics of the linguistic patterns in participant generated descriptions of these thought types. We show how, by positioning language as a window into spontaneous cognition, existing theories on these attentional states can be updated and reaffirmed. Our findings align with prior research, reinforcing that D ej ` a Vu is a metacognitive experience characterized by abstract and spatial language, Involuntary Autobiographical Memories are rich in personal and emotionally significant detail, and Unexpected Thoughts are marked by unpredictability and cognitive disruption. This work is demonstrative of languages' potential to reveal deeper insights into how internal spontaneous cognitive states manifest through expression.

Scientific text is complex as it contains technical terms by definition. Simplifying such text for non-domain experts enhances accessibility of innovation and information. Politicians could be enabled to understand new findings on topics on which they intend to pass a law, or family members of seriously ill patients could read about clinical trials. The SimpleText CLEF Lab focuses on exactly this problem of simplification of scientific text. Task 1.1 of the 2025 edition specifically handles the simplification of complex sentences, so very short texts with little context. To tackle this task we investigate the identification of complex terms in sentences which are rephrased using small Gemini and OpenAI large language models for non-expert readers.

This paper presents our submission to the ACMMM25 - Grand Challenge on Multimedia Verification. We developed a multi-agent verification system that combines Multimodal Large Language Models (MLLMs) with specialized verification tools to detect multimedia misinformation. Our system operates through six stages: raw data processing, planning, information extraction, deep research, evidence collection, and report generation. The core Deep Researcher Agent employs four tools: reverse image search, metadata analysis, fact-checking databases, and verified news processing that extracts spatial, temporal, attribution, and motivational context. We demonstrate our approach on a challenge dataset sample involving complex multimedia content. Our system successfully verified content authenticity, extracted precise geolocation and timing information, and traced source attribution across multiple platforms, effectively addressing real-world multimedia verification scenarios.

Machine learning methods are increasingly applied to analyze health-related public discourse based on large-scale data, but questions remain regarding their ability to accurately detect different types of health sentiments. Especially, Large Language Models (LLMs) have gained attention as a powerful technology, yet their accuracy and feasibility in capturing different opinions and perspectives on health issues are largely unexplored. Thus, this research examines how accurate the three prominent LLMs (GPT, Gemini, and LLAMA) are in detecting risk-promoting versus health-supporting sentiments across two critical public health topics: Human Papillomavirus (HPV) vaccination and heated tobacco products (HTPs). Drawing on data from Facebook and Twitter, we curated multiple sets of messages supporting or opposing recommended health behaviors, supplemented with human annotations as the gold standard for sentiment classification. The findings indicate that all three LLMs generally demonstrate substantial accuracy in classifying risk-promoting and health-supporting sentiments, although notable discrepancies emerge by platform, health issue, and model type. Specifically, models often show higher accuracy for risk-promoting sentiment on Facebook, whereas health-supporting messages on Twitter are more accurately detected. An additional analysis also shows the challenges LLMs face in reliably detecting neutral messages. These results highlight the importance of carefully selecting and validating language models for public health analyses, particularly given potential biases in training data that may lead LLMs to overestimate or underestimate the prevalence of certain perspectives.

Visual question answering (VQA) in medical imaging aims to support clinical diagnosis by automatically interpreting complex imaging data in response to natural language queries. Existing studies typically rely on distinct visual and textual encoders to independently extract features from medical images and clinical questions, which are subsequently combined to generate answers. Specifically, in computed tomography (CT), such approaches are similar to the conventional practices in medical image analysis. However, these approaches pay less attention to the spatial continuity and inter-slice correlations in the volumetric CT data, leading to fragmented and imprecise responses. In this paper, we propose a novel large language model (LLM)-based framework enhanced by a graph representation of salient features. Different from conventional multimodal encoding strategies, our approach constructs a cross-modal graph integrating both visual and textual features, treating individual CT slices and question tokens as nodes within the graph. We further leverage an attentive graph convolutional network to dynamically fuse information within this structure. The resulting aggregated graph features then serve as a soft prompt to guide a large language model in generating accurate answers. Extensive experiments on the M3D-VQA benchmark demonstrate that our approach consistently outperforms baselines across multiple evaluation metrics, offering more robust reasoning capabilities.