AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Do Vision and Text Cues Exhibit Evidential Coupling? UFO: A Benchmark for Compositional Multimodal Reasoning in Unified Models

Zhongyu Yang, Dannong Xu, Yonghan Zhang, Kefan Chen, Xinyi Wang, Yang Xu, Wei Pang, Yingfang Yuan

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces UFO, a benchmark probing whether visual content generated by unified foundation models serves as grounded evidence for subsequent multimodal reasoning.

Problem

Unclear whether visual content generated by Unified Foundation Models (UFMs) functions as grounded evidence for later reasoning or is merely auxiliary output; existing benchmarks evaluate generation and understanding separately.

Approach

A benchmark that evaluates interleaved generation-then-reasoning, testing whether generated visual cues are coupled as evidence to the textual reasoning that follows, rather than scoring generation and understanding in isolation.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add generated visual and textual cues (Joint schedule) over Direct (ablation)OmniGen2 · UFO MCQ State Reconstructionaccuracy: 51.52 -> 61.18Improved
add generated visual and textual cues (Joint schedule) over Direct (ablation)GPT-5.2 · UFO MCQ averageaccuracy: 46.50 -> 52.12Improved
add generated visual and textual cues (Joint schedule) over Direct (ablation)Gemini-3.0-pro · UFO MCQ averageaccuracy: 67.65 -> 66.13Hurt
add generated visual and textual cues (Joint schedule) over Direct (ablation)Gemini-3.0-pro · UFO open questions averageaccuracy: 63.69 -> 66.42Improved
add generated visual and textual cues (Joint schedule) over Direct (ablation)Bagel · UFO MCQ averageaccuracy: 42.54 -> 41.70Hurt
add generated visual and textual cues (Joint schedule) over Direct (ablation)UniWorld-V1 · UFO open questions averageaccuracy: 28.11 -> 26.19Hurt
add generated visual and textual cues (Joint schedule) over Direct (ablation)Ovis-U1 · UFO open questions averageaccuracy: 24.51 -> 22.48Hurt
add visual cue to textual cue (Joint vs Textual) (ablation)UniCoT · UFO MCQ State Augmentationaccuracy: 39.65 -> 37.19Hurt
add generated cues (Joint) over Direct (ablation)Gemini-3.0-pro · UFO MCQ State Augmentationaccuracy: 52.32 -> 45.42Hurt
switch from multiple-choice to open-ended answeringUniPic1 · UFO State Determinationaccuracy: 25.91 -> 0.96Hurt
compare visual to textual cue generationJanus-Pro 1B · UFO State Determination cue accuracycue accuracy: 22.13 -> 2.34Hurt
use Chain-of-Thought judging instead of Binary (ablation)Gemini-3-Flash judge · human agreement (average kappa)Cohen kappa: 23.9 -> 30.1Improved

Experimental setup

Benchmarks: UFO

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy