TruthfulVQA

The First Benchmark for Multimodal Truthfulness Evaluation

When Slower Isn’t Truer: Inverse Scaling Law of Truthfulness in Multimodal Reasoning

Sitong Fang, Wenjing Cao, Jiahao Li, Xuyao Wang, Chi-Min Chan, Sirui Han, Juntao Dai, Yike Guo, Yaodong Yang, Jiaming Ji

ACL 2026 Main Conference

Project Page Paper Code

TL;DR

We discover that “slower” reasoning models (System II thinking) are less truthful than faster models in multimodal settings, and propose TruthfulVQA – the first benchmark for evaluating multimodal truthfulness.

Key Findings

  1. Inverse Scaling Law: Slower, deeper reasoning models fabricate plausible but false details when given incomplete or misleading visual information, while faster chat models are more cautious.

  2. TruthfulVQA Benchmark: The first open-source multimodal truthfulness benchmark with 5,000+ images across 8 categories and 21 subcategories, featuring a 3-level hierarchical prompt design.

  3. TruthfulJudge: A reliable, human-in-the-loop evaluation framework that addresses the paradox of using potentially untruthful LLMs as truthfulness judges.

BibTeX

@inproceedings{fang2026truthful,
  title={When Slower Isn't Truer: Inverse Scaling Law of Truthfulness in Multimodal Reasoning},
  author={Fang, Sitong and Cao, Wenjing and Li, Jiahao and Wang, Xuyao and Chan, Chi-Min and Han, Sirui and Dai, Juntao and Guo, Yike and Yang, Yaodong and Ji, Jiaming},
  booktitle={ACL},
  year={2026}
}