How Do Language Models Acquire Character-Level Information?

Language models (LMs) have been reported to implicitly encode character-level information, despite not being explicitly provided during training. However, the mechanisms underlying this phenomenon remain largely unexplored. To reveal the mechanisms, we analyze how models acquire character-level knowledge by comparing LMs trained under controlled settings, such as specifying the pre-training dataset or tokenizer, with those trained under standard settings. We categorize the contributing factors into those independent of tokenization. Our analysis reveals that merge rules and orthographic constraints constitute primary factors arising from tokenization, whereas semantic associations of substrings and syntactic information function as key factors independent of tokenization.

JSynFlow: Japanese Synthesised Flowchart Visual Question Answering Dataset built with Large Language Models

Vision and language models (VLMs) are expected to analyse complex documents, such as those containing flowcharts, through a question-answering (QA) interface. The ability to recognise and interpret these flowcharts is in high demand, as they provide valuable insights unavailable in text-only explanations. However, developing VLMs with precise flowchart understanding requires large-scale datasets of flowchart images and corresponding text, the creation of which is highly time-consuming. To address this challenge, we introduce JSynFlow, a synthesised visual QA dataset for Japanese flowcharts, generated using large language models (LLMs). Our dataset comprises task descriptions for various business occupations, the corresponding flowchart images rendered from domain-specific language (DSL) code, and related QA pairs. This paper details the dataset's synthesis procedure and demonstrates that fine-tuning with JSynFlow significantly improves VLM performance on flowchart-based QA tasks. Our dataset is publicly available at https://huggingface.co/datasets/jri-advtechlab/jsynflow.

Can vision language models learn intuitive physics from interaction?

Pre-trained vision language models do not have good intuitions about the physical world. Recent work has shown that supervised fine-tuning can improve model performance on simple physical tasks. However, fine-tuned models do not appear to learn robust physical rules that can generalize to new contexts. Based on research in cognitive science, we hypothesize that models need to interact with an environment to properly learn its physical dynamics. We train models that learn through interaction with the environment using reinforcement learning. While learning from interaction allows models to improve their within-task performance, it fails to produce models with generalizable physical intuitions. We find that models trained on one task do not reliably generalize to related tasks, even if the tasks share visual statistics and physical principles, and regardless of whether the models are trained through interaction.

Hallucination-Resistant Security Planning with a Large Language Model

Large language models (LLMs) are promising tools for supporting security management tasks, such as incident response planning. However, their unreliability and tendency to hallucinate remain significant challenges. In this paper, we address these challenges by introducing a principled framework for using an LLM as decision support in security management. Our framework integrates the LLM in an iterative loop where it generates candidate actions that are checked for consistency with system constraints and lookahead predictions. When consistency is low, we abstain from the generated actions and instead collect external feedback, e.g., by evaluating actions in a digital twin. This feedback is then used to refine the candidate actions through in-context learning (ICL). We prove that this design allows to control the hallucination risk by tuning the consistency threshold. Moreover, we establish a bound on the regret of ICL under certain assumptions. To evaluate our framework, we apply it to an incident response use case where the goal is to generate a response and recovery plan based on system logs. Experiments on four public datasets show that our framework reduces recovery times by up to 30% compared to frontier LLMs.

Once Correct, Still Wrong: Counterfactual Hallucination in Multilingual Vision-Language Models

Vision-language models (VLMs) can achieve high accuracy while still accepting culturally plausible but visually incorrect interpretations. Existing hallucination benchmarks rarely test this failure mode, particularly outside Western contexts and English. We introduce M2CQA, a culturally grounded multimodal benchmark built from images spanning 17 MENA countries, paired with contrastive true and counterfactual statements in English, Arabic, and its dialects. To isolate hallucination beyond raw accuracy, we propose the CounterFactual Hallucination Rate (CFHR), which measures counterfactual acceptance conditioned on correctly answering the true statement. Evaluating state-of-the-art VLMs under multiple prompting strategies, we find that CFHR rises sharply in Arabic, especially in dialects, even when true-statement accuracy remains high. Moreover, reasoning-first prompting consistently increases counterfactual hallucination, while answering before justifying improves robustness. We will make the experimental resources and dataset publicly available for the community.

Cross-Lingual Empirical Evaluation of Large Language Models for Arabic Medical Tasks

In recent years, Large Language Models (LLMs) have become widely used in medical applications, such as clinical decision support, medical education, and medical question answering. Yet, these models are often English-centric, limiting their robustness and reliability for linguistically diverse communities. Recent work has highlighted discrepancies in performance in low-resource languages for various medical tasks, but the underlying causes remain poorly understood. In this study, we conduct a cross-lingual empirical analysis of LLM performance on Arabic and English medical question and answering. Our findings reveal a persistent language-driven performance gap that intensifies with increasing task complexity. Tokenization analysis exposes structural fragmentation in Arabic medical text, while reliability analysis suggests that model-reported confidence and explanations exhibit limited correlation with correctness. Together, these findings underscore the need for language-aware design and evaluation strategies in LLMs for medical tasks.

Transport and Merge: Cross-Architecture Merging for Large Language Models

Large language models (LLMs) achieve strong capabilities by scaling model capacity and training data, yet many real-world deployments rely on smaller models trained or adapted from low-resource data. This gap motivates the need for mechanisms to transfer knowledge from large, high-resource models to smaller, low-resource targets. While model merging provides an effective transfer mechanism, most existing approaches assume architecture-compatible models and therefore cannot directly transfer knowledge from large high-resource LLMs to heterogeneous low-resource targets. In this work, we propose a cross-architecture merging framework based on optimal transport (OT) that aligns activations to infer cross-neuron correspondences between heterogeneous models. The resulting transport plans are then used to guide direct weight-space fusion, enabling effective high-resource to low-resource transfer using only a small set of inputs. Extensive experiments across low-resource languages and specialized domains demonstrate consistent improvements over target models.

SciDef: Automating Definition Extraction from Academic Literature with Large Language Models

Definitions are the foundation for any scientific work, but with a significant increase in publication numbers, gathering definitions relevant to any keyword has become challenging. We therefore introduce SciDef, an LLM-based pipeline for automated definition extraction. We test SciDef on DefExtra & DefSim, novel datasets of human-extracted definitions and definition-pairs' similarity, respectively. Evaluating 16 language models across prompting strategies, we demonstrate that multi-step and DSPy-optimized prompting improve extraction performance. To evaluate extraction, we test various metrics and show that an NLI-based method yields the most reliable results. We show that LLMs are largely able to extract definitions from scientific literature (86.4% of definitions from our test-set); yet future work should focus not just on finding definitions, but on identifying relevant ones, as models tend to over-generate them. Code & datasets are available at https://github.com/Media-Bias-Group/SciDef.

DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders

Sparse autoencoders (SAEs) have become a standard tool for mechanistic interpretability in autoregressive large language models (LLMs), enabling researchers to extract sparse, human-interpretable features and intervene on model behavior. Recently, as diffusion language models (DLMs) have become an increasingly promising alternative to the autoregressive LLMs, it is essential to develop tailored mechanistic interpretability tools for this emerging class of models. In this work, we present DLM-Scope, the first SAE-based interpretability framework for DLMs, and demonstrate that trained Top-K SAEs can faithfully extract interpretable features. Notably, we find that inserting SAEs affects DLMs differently than autoregressive LLMs: while SAE insertion in LLMs typically incurs a loss penalty, in DLMs it can reduce cross-entropy loss when applied to early layers, a phenomenon absent or markedly weaker in LLMs. Additionally, SAE features in DLMs enable more effective diffusion-time interventions, often outperforming LLM steering. Moreover, we pioneer certain new SAE-based research directions for DLMs: we show that SAEs can provide useful signals for DLM decoding order; and the SAE features are stable during the post-training phase of DLMs. Our work establishes a foundation for mechanistic interpretability in DLMs and shows a great potential of applying SAEs to DLM-related tasks and algorithms.

CASTLE: A Comprehensive Benchmark for Evaluating Student-Tailored Personalized Safety in Large Language Models

Large language models (LLMs) have advanced the development of personalized learning in education. However, their inherent generation mechanisms often produce homogeneous responses to identical prompts. This one-size-fits-all mechanism overlooks the substantial heterogeneity in students cognitive and psychological, thereby posing potential safety risks to vulnerable groups. Existing safety evaluations primarily rely on context-independent metrics such as factual accuracy, bias, or toxicity, which fail to capture the divergent harms that the same response might cause across different student attributes. To address this gap, we propose the concept of Student-Tailored Personalized Safety and construct CASTLE based on educational theories. This benchmark covers 15 educational safety risks and 14 student attributes, comprising 92,908 bilingual scenarios. We further design three evaluation metrics: Risk Sensitivity, measuring the model ability to detect risks; Emotional Empathy, evaluating the model capacity to recognize student states; and Student Alignment, assessing the match between model responses and student attributes. Experiments on 18 SOTA LLMs demonstrate that CASTLE poses a significant challenge: all models scored below an average safety rating of 2.3 out of 5, indicating substantial deficiencies in personalized safety assurance.