More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models

TL;DR

This paper investigates the tendency of multimodal reasoning models to hallucinate and drift from visual grounding during extended reasoning, introducing metrics and benchmarks to evaluate and understand this trade-off.

Contribution

It introduces RH-AUC and RH-Bench, new tools for systematically measuring the balance between reasoning ability and visual perception in multimodal models.

Findings

Larger models better balance reasoning and perception.

Training data types and domains influence this balance more than data volume.

Longer reasoning chains reduce focus on visual inputs.

Abstract

Test-time compute has empowered multimodal large language models to generate extended reasoning chains, yielding strong performance on tasks such as multimodal math reasoning. However, this improved reasoning ability often comes with increased hallucination: as generations become longer, models tend to drift away from image-grounded content and rely more heavily on language priors. Attention analysis shows that longer reasoning chains lead to reduced focus on visual inputs, which contributes to hallucination. To systematically study this phenomenon, we introduce RH-AUC, a metric that quantifies how a model's perception accuracy changes with reasoning length, allowing us to evaluate whether the model preserves visual grounding during reasoning. We also release RH-Bench, a diagnostic benchmark that spans a variety of multimodal tasks, designed to assess the trade-off between reasoning ability and hallucination. Our analysis reveals that (i) larger models typically achieve a better balance between reasoning and perception, and (ii) this balance is influenced more by the types and domains of training data than by its overall volume. These findings underscore the importance of evaluation frameworks that jointly consider both reasoning quality and perceptual fidelity.