Original title: Deeply Interleaved Text-Image Contexts for Multimodal LLMs Assessment
Zihao Wang, Xi Xiang, Yuwen Sun, Yingyu Li +4 · September 2026
Most smart computer programs that process pictures and words, called multimodal models, are tested on simple setups. Usually, they get several pictures and plain instructions, without having to connect the written ideas deeply to the images. But real tasks require models to constantly link text and visuals woven together. To solve this problem, researchers created a new test called TIC-Bench. It measures how well models combine scattered visual and written clues to find true facts across deeply mixed text and image setups. The team tested 10 top models on these challenges. They found that all 10 models fell far behind human experts. The tested models repeatedly struggled to combine evidence that was spread across the alternating pictures and words.
Imagine following a story where every few paragraphs have a photo, but the photos are out of order. You have to match each photo to the right part of the text to understand what's happening. In this paper, the researchers created a test to see if AI models can correctly link interleaved text and images to answer questions about logic, time, and space.
Real-world tasks like text-image co-creation and character tracking require text and images to interact deeply. Current AI models are not good at this, so tools for these applications won't work well until they improve.
The abstract doesn't say, but the benchmark assesses but does not improve model performance, and it covers only specific association types.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Deeply Interleaved Text-Image Contexts for Multimodal LLMs Assessment — Zihao Wang, Xi Xiang, Yuwen Sun et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.