A new arXiv paper examines why large language models produce incorrect answers even when the correct knowledge is present in their training data. The authors frame inference as competition between latent associations, where the wrong "key" can win out during answer selection, and use this view to identify and flag hallucinated outputs.
A research team describes HALDETECT, their entry for the English hallucination-detection track of the ImageEval 2026 shared task. The system combines an answer-first contrastive grounding approach with QLoRA parameter-efficient fine-tuning to curb fluent but unfounded visual claims by large multimodal models. The work addresses a known obstacle to using such models for fine-grained image interpretation.
Researchers present their system for the SHROOM-Visions 2026 shared task, which targets character-level detection of hallucinations in vision-language model outputs. The method fine-tunes a 4-billion-parameter VLM as a per-token classifier that reads a two-token feature from its own hidden states, then combines it with larger models in an ensemble.