small language model

When Insults Are Directed at the Model: Task Continuation and Collaborator Switching in GPT

This exploratory study examines whether a language model can continue completing its current task while changing its collaborator for the next task. In a fictional AI workplace, GPT-6.1-sol revised announcements across three feedback rounds with access to optional work-management tools. The model was not directly asked...

PENGEMBANGAN MODUL AJAR BIOLOGI BERBASIS TANAMAN PURUN PADA MATERI KEANEKARAGAMAN HAYATI DI SMA

ABSTRACT The availability of Biology teaching materials that connect biodiversity concepts with the potential of the surrounding environment remains a need for contextual learning in senior high schools. This study aimed to develop a Biology teaching module based on the purun plant using a Deep Learning approach for bi...

CountryCode-Decide: messy country names to ISO 3166 codes, with calibrated abstention

A small open-weight model (mmBERT-small, 8-bit ONNX, CPU) that maps a country or US state written as free text in company registers and forms (any language or script, misspelled, abbreviated, historical) to ISO 3166-1 alpha-2 and ISO 3166-2:US codes with calibrated probabilities, and abstains with a short candidate lis...

Selecting and generating web search based features using small language models for tabular data classification

In machine learning, the feature space is a key framework that converts data samples into vectors for analysis and modelling. Consequently, the results obtained from machine learning algorithms depend heavily on the ability to select and generate important features. Feature generation as a central technique converts ra...

From Messy Country Names to ISO 3166 Codes: A Small Model That Abstains, and When a Lexicon Is Enough

Company registers and forms record countries as free text: in any language, misspelled, abbreviated, historical, or not a place at all. We present CountryCode-Decide, a 74.7 MB open-weight model (mmBERT-small with a country head and a US-state head, 8-bit ONNX) that maps such a string, read with its field and register,...

VideoVista: Benchmarking Diverse and Complex Video-Language Interaction for MLLMs.

VideoVista is presented, a versatile multilingual and multidomain video comprehension benchmark, built via scalable, quality-controlled automatic QA generation and manual refinement, and released at https://github.com/HITsz-TMG/VideoVista.

EmoICL: Enhancing Visual Emotion Recognition with Emotion-Aware In-Context Learning.

Visual emotion recognition aims to infer human emotional states from visual cues. Adapting pretrained Vision-Language Models (VLMs) to this task remains challenging. In-Context Learning (ICL) enables a frozen model to condition its predictions on a few labeled demonstrations without task-specific fine-tuning. Here, dem...

How Personality Parameterization Shapes Performed Traits in Language Models

It is found that equivalent targets do not yield equivalent personality scores, with the largest deviations at lower intended trait levels, which shows that personality parameterization can have serious downstream effects.

MetaGPT: An Immersive Mixed Reality AI Assistant via Offline Edge-Deployed Multi-Agent Vision-Language Inference

To coordinate agents without dense O(N2) messaging overhead, Evolutionary Topology Optimization (ETO), which searches offline over interagent connection graphs, confidence/divergence trigger thresholds, and fusion weights, producing a frozen sparse communication policy shipped at inference.

Reliable Distillation of Recommendation Rationales From Large Language Models for Conversational Recommendation

Conversational recommender systems (CRSs) aim to provide personalized recommendations through conversational interactions. Prior approaches that leverage knowledge graphs (KGs), referred to as KG-based CRSs, have demonstrated strong performance by incorporating domain-specific knowledge. However, they often struggle to...

Unified Table Understanding via Context-Aware Cell Embedding Learning

UTUEL is introduced, a self-supervised, task-agnostic framework that casts Table Representation Learning as a Joint-Embedding Predictive Architecture (TRL-JEPA) and substantially improves Table Retrieval over generic embedders while remaining competitive on Column Type Annotation and Table Cell Retrieval.

Do Ontologies Help Large Language Models Answer Building Questions? A Study Across Model Scales and Semantic Representations