Sentences with multiple quantifiers often lead to interpretive ambiguities, which can vary across languages. This study adopts a cross-linguistic approach to examine how large language models (LLMs) handle quantifier scope interpretation in English and Chinese, using probabilities to assess interpretive likelihood. Human similarity (HS) scores were used to quantify the extent to which LLMs emulate human performance across language groups. Results reveal that most LLMs prefer the surface scope interpretations, aligning with human tendencies, while only some differentiate between English and Chinese in the inverse scope preferences, reflecting human-similar patterns. HS scores highlight variability in LLMs' approximation of human behavior, but their overall potential to align with humans is notable. Differences in model architecture, scale, and particularly models' pre-training data language background, significantly influence how closely LLMs approximate human quantifier scope interpretations.

Effective long-term memory in conversational AI requires synthesizing information across multiple sessions. However, current systems place excessive reasoning burden on response generation, making performance significantly dependent on model sizes. We introduce PREMem (Pre-storage Reasoning for Episodic Memory), a novel approach that shifts complex reasoning processes from inference to memory construction. PREMem extracts fine-grained memory fragments categorized into factual, experiential, and subjective information; it then establishes explicit relationships between memory items across sessions, capturing evolution patterns like extensions, transformations, and implications. By performing this reasoning during pre-storage rather than when generating a response, PREMem creates enriched representations while reducing computational demands during interactions. Experiments show significant performance improvements across all model sizes, with smaller models achieving results comparable to much larger baselines while maintaining effectiveness even with constrained token budgets. Code and dataset are available at https://github.com/sangyeop-kim/PREMem.

Yixuan T ang Yi Y ang The Hong Kong University of Science and Technology ytangch@connect.ust.hk, Domain-specific embedding models have shown promise for applications that require specialized semantic understanding, such as coding agents and financial retrieval systems, often achieving higher performance gains than general models. However, state-of-the-art embedding models are typically based on LLMs, which contain billions of parameters, making deployment challenging in resource-constrained environments. Model compression through pruning offers a promising solution, but existing pruning methods treat all parameters uniformly, failing to distinguish between general semantic representations and domain-specific patterns, leading to suboptimal pruning decisions. Thus, we propose GAPrune, a pruning framework that addresses this challenge by considering both domain importance and preserving general linguistic foundation. Our method uses Fisher Information to measure importance and general-domain gradient alignment to assess parameter behavior, then combines these signals using our Domain Alignment Importance (DAI) scoring. Lower DAI scores indicate that the parameter is either less important for the domain task or creates conflicts between domain and general objectives. Experiments on two domain benchmarks, FinMTEB and ChemTEB, show that GAPrune maintains performance within 2.5% of dense models in one-shot pruning at 50% sparsity, while outperforming all baselines. With retraining in 100 steps, GAPrune achieves +4.51% improvement on FinMTEB and +1.73% on ChemTEB, demonstrating that our pruning strategy not only preserves but enhances domain-specific capabilities. The deployment of large language models in specialized domains has revealed a critical challenge: while general-purpose models excel at broad language understanding, they often fail to capture domain-specific semantics crucial for real-world applications (Gu et al., 2021; Y ao et al., 2024). This semantic gap is evident for embedding models, where precise representation of domain-specific concepts directly impacts downstream task performance.

Large Language Models (LLMs) have demonstrated substantial progress in biomedical and clinical applications, motivating rigorous evaluation of their ability to answer nuanced, evidence-based questions. We curate a multi-source benchmark drawing from Cochrane systematic reviews and clinical guidelines, including structured recommendations from the American Heart Association and narrative guidance used by insurers. Using GPT-4o-mini and GPT-5, we observe consistent performance patterns across sources and clinical domains: accuracy is highest on structured guideline recommendations (90%) and lower on narrative guideline and systematic review questions (60--70%). We also find a strong correlation between accuracy and the citation count of the underlying systematic reviews, where each doubling of citations is associated with roughly a 30% increase in the odds of a correct answer. Models show moderate ability to reason about evidence quality when contextual information is supplied. When we incorporate retrieval-augmented prompting, providing the gold-source abstract raises accuracy on previously incorrect items to 0.79; providing top 3 PubMed abstracts (ranked by semantic relevance) improves accuracy to 0.23, while random abstracts reduce accuracy (0.10, within temperature variation). These effects are mirrored in GPT-4o-mini, underscoring that source clarity and targeted retrieval -- not just model size -- drive performance. Overall, our results highlight both the promise and current limitations of LLMs for evidence-based clinical question answering. Retrieval-augmented prompting emerges as a useful strategy to improve factual accuracy and alignment with source evidence, while stratified evaluation by specialty and question type remains essential to understand current knowledge access and to contextualize model performance.

Although LLM-based conversational agents demonstrate strong fluency and coherence, they still produce undesirable behaviors (errors) that are challenging to prevent from reaching users during deployment. Recent research leverages large language models (LLMs) to detect errors and guide response-generation models toward improvement. However, current LLMs struggle to identify errors not explicitly specified in their instructions, such as those arising from updates to the response-generation model or shifts in user behavior. In this work, we introduce Automated Error Discovery, a framework for detecting and defining errors in conversational AI, and propose SEEED (Soft Clustering Extended Encoder-Based Error Detection), as an encoder-based approach to its implementation. We enhance the Soft Nearest Neighbor Loss by amplifying distance weighting for negative samples and introduce Label-Based Sample Ranking to select highly contrastive examples for better representation learning. SEEED outperforms adapted baselines -- including GPT-4o and Phi-4 -- across multiple error-annotated dialogue datasets, improving the accuracy for detecting unknown errors by up to 8 points and demonstrating strong generalization to unknown intent detection.

Transformers have become the foundation for a wide range of state--of--the--art models across natural language processing, computer vision, and other machine learning domains. Despite their widespread deployment, the robustness of these models under fault conditions remains underexplored. We present GoldenTransformer, a modular and extensible fault injection framework designed to evaluate the resiliency of Large Language Models to induced hardware faults. GoldenTransformer offers a unified Python-based platform for injecting diverse classes of faults--such as weight corruption, activation injections, and attention--level disruptions--into pretrained transformer--based models. Inspired by the GoldenEye simulator for DNNs, our framework focuses on the unique challenges of working with large transformer architectures, including considerations such as structural complexity, latent dependencies, and nonuniform layer definitions. GoldenTransformer is built atop PyTorch and HuggingFace Transformers, and it supports experiment reproducibility, metric logging, and visualization out of the box. We detail the technical design and use of GoldenTransformer and demonstrate through several example experiments on classification and generation tasks. By enabling controlled injection of faults at multiple logical and structural points in a transformer, GoldenTransformer offers researchers and practitioners a valuable tool for model robustness analysis and for guiding dependable system design in real-world LLM applications.

Generative AI (GenAI) is rapidly entering K-12 classrooms, offering teachers new ways for teaching practices. Yet GenAI models are often trained on culturally uneven datasets, embedding a "default culture" that often misaligns with local classrooms. To understand how teachers navigate this gap, we defined the new concept Cultural Distance (the gap between GenAI's default cultural repertoire and the situated demands of teaching practice) and conducted in-depth interviews with 30 K-12 teachers, 10 each from South Africa, Taiwan, and the United States, who had integrated AI into their teaching practice. These teachers' experiences informed the development of our three-level cultural distance framework. This work contributes the concept and framework of cultural distance, six illustrative instances spanning in low, mid, high distance levels with teachers' experiences and strategies for addressing them. Empirically, we offer implications to help AI designers, policymakers, and educators create more equitable and culturally responsive GenAI tools for education.

AI answer engines increasingly mediate access to domain knowledge by generating responses and citing web sources. We introduce GEO-16, a 16 pillar auditing framework that converts on page quality signals into banded pillar scores and a normalized GEO score G that ranges from 0 to 1. Using 70 product intent prompts, we collected 1,702 citations across three engines (Brave Summary, Google AI Overviews, and Perplexity) and audited 1,100 unique URLs. In our corpus, the engines differed in the GEO quality of the pages they cited, and pillars related to Metadata and Freshness, Semantic HTML, and Structured Data showed the strongest associations with citation. Logistic models with domain clustered standard errors indicate that overall page quality is a strong predictor of citation, and simple operating points (for example, G at least 0.70 combined with at least 12 pillar hits) align with substantially higher citation rates in our data. We report per engine contrasts, vertical effects, threshold analysis, and diagnostics, then translate findings into a practical playbook for publishers. The study is observational and focuses on English language B2B SaaS pages; we discuss limitations, threats to validity, and reproducibility considerations.

Large Language Models (LLMs) exhibit impressive reasoning and question-answering capabilities. However, they often produce inaccurate or unreliable content known as hallucinations. This unreliability significantly limits their deployment in high-stakes applications. Thus, there is a growing need for a general-purpose method to detect hallucinations in LLMs. In this work, we introduce HalluField, a novel field-theoretic approach for hallucination detection based on a parametrized variational principle and thermodynamics. Inspired by thermodynamics, HalluField models an LLM's response to a given query and temperature setting as a collection of discrete likelihood token paths, each associated with a corresponding energy and entropy. By analyzing how energy and entropy distributions vary across token paths under changes in temperature and likelihood, HalluField quantifies the semantic stability of a response. Hallucinations are then detected by identifying unstable or erratic behavior in this energy landscape. HalluField is computationally efficient and highly practical: it operates directly on the model's output logits without requiring fine-tuning or auxiliary neural networks. Notably, the method is grounded in a principled physical interpretation, drawing analogies to the first law of thermodynamics. Remarkably, by modeling LLM behavior through this physical lens, HalluField achieves state-of-the-art hallucination detection performance across models and datasets.

As scientific knowledge grows at an unprecedented pace, evaluation benchmarks must evolve to reflect new discoveries and ensure language models are tested on current, diverse literature. We propose a scalable, modular framework for generating multiple-choice question-answering (MCQA) benchmarks directly from large corpora of scientific papers. Our pipeline automates every stage of MCQA creation, including PDF parsing, semantic chunking, question generation, and model evaluation. As a case study, we generate more than 16,000 MCQs from 22,000 open-access articles in radiation and cancer biology. We then evaluate a suite of small language models (1.1B-14B parameters) on these questions, comparing baseline accuracy with retrieval-augmented generation (RAG) from paper-derived semantic chunks and from reasoning traces distilled from GPT-4.1. We find that reasoning-trace retrieval consistently improves performance on both synthetic and expert-annotated benchmarks, enabling several small models to surpass GPT-4 on the 2023 Astro Radiation and Cancer Biology exam.