small language model

How Fragile Is On-Device Language Model Safety? Localizing Safety-Critical Parameters for Sparse Fault Analysis

As small language models (SLMs) are increasingly deployed on resource-constrained and on-device platforms, including as components of agentic systems, the integrity of locally stored model parameters becomes an important safety concern. We investigate whether safety-sensitive behavior in LLaMA-2-7B-Chat is concentrated...

From Information to Control

This preprint develops an end-to-end causal account of how information acquires behavioral control in language models. Across a coordinated series of mechanistic studies, we trace the computation from distributed relational and typed representations, through task- and recipient-dependent qualification, contextual multi...

Trust in AI-Generated Software: Why verification, not generation, is the new bottleneck – a European perspective

Large language models now write a substantial share of new source code, yet the mechanisms by which organisations establish confidence in that code have hardly changed. Independent testing indicates that AI-generated code introduces security weaknesses in close to half of evaluated tasks, while European regulation – no...

A Lightweight LINE-Based Knowledge Chatbot for Mobile Technical Certification Learning: An Empirical Study with ChatGPT as a Baseline

This study developed a lightweight LINE-based mobile knowledge chatbot for the Level B Computer Hardware Repair certification subject and examined its instructional effectiveness in comparison with ChatGPT. The system was deployed through the LINE mobile messaging platform, allowing learners to access certification-rel...

PRISM-LoRA: Principal-Direction-Guided Single-LoRA Merging Framework for Continual Learning of Large Language Models

PRISM-LoRA is proposed, a replay-free framework that repeatedly trains a single LoRA module, merges its update into the backbone, and reinitializes it for the next task, unlike LoRA-based continual-learning approaches that retain task-specific modules or allocate separate task-specific learning spaces.

Prompt to Bench: language models as a no-CAD entry point to 3D printing for biology labs

A living paper, student tutorial, library of 16 tested parametric OpenSCAD designs for biology labs (benchware, tools, quick fixes and parts of full instruments), and Prompt-to-Bench-16, a benchmark that scores language-model-written OpenSCAD against hidden geometric checks. Includes an evaluation of small open-weight...

NeST: Neuron Selective Tuning for LLM Safety

NeST: Neuron-Selective Tuning for Efficient Safety Alignment A lightweight and structure-aware safety alignment framework that selectively adapts safety-relevant neurons to strengthen refusal behavior while preserving the model's general capabilities. 🚀 Overview Safety alignment is essential for reliable deployment of...

Associations between content and engagement vary across social media communities

Research on social media engagement often seeks general content features that are associated with online success. However, social media platforms contain communities with different audiences, topics, and norms, and the same features may have different associations with engagement within different communities. This rese...

Safety, Accuracy, Empathy, Information Quality, and Readability of Large Language Models in Answering Caregiver Questions About Pediatric Burn First Aid and Inpatient Care: A Cross-Sectional Comparative Study

Machines Can Produce What We Speak in the Brain: A Linguistic Framework for Decoding Inner Speech

When we rehearse a sentence silently, such as ‘Hi, what’s up?’ before greeting a friend, the brain runs much of the machinery of speaking without moving a muscle. This paper asks, from a linguistic perspective, how a machine could turn that silent speech into words. We propose a framework in two parts. The first is a m...

A Dataset for Multidimensional Evaluation of French Synthetic Texts Generated by SLMs

This dataset comprises synthetic news descriptions in French, generated from real newspaper headlines using a controlled Small Language Models (SLM) pipeline. The generation process follows three distinct configurations: Retrieval-Augmented Generation (RAG) and generation without external context (NO RAG) and QLoRA-bas...

Reporting adherence of meridian-sinew Tuina for non-specific neck pain to a 2026 expert consensus

This registration is retrospective. It records a systematic appraisal that was completed between 24 September and 1 November 2026. The protocol, the search strategy and the eligibility criteria were finalised in writing before data extraction began; the appraisal itself, however, was finished before this registration w...