VERIFY
A Benchmark of Visual Explanation and Reasoning
for Investigating Multimodal Reasoning FidelitY

Jing Bi*, JunJia Guo*, Susan Liang*, Guangyu Sun†, Luchuan Song*, Yunlong Tang*, Jinxi He*, Jiarui Wu*, Ali Vosoughi*, Chen Chen†, Chenliang Xu*

*University of Rochester, †UCF

We introduce VERIFY, the first benchmark explicitly designed to assess the reasoning path of MLLMs in visual reasoning tasks. We propose novel metrics that assess reasoning fidelity beyond accuracy. Our evaluation reveals critical limitations in current MLLMs, emphasizing the need for a more balanced approach to perception and reasoning.

Teaser

O1: 80 second, still wrong

In the color XOR reasoning challenge, O1 takes 1 minute and 20 seconds to propose a solution that ultimately proves incorrect (answer is B). Note: the video plays at normal speed.

I. Data

Categories from the VERIFY dataset cover a range of patterns, from logical operations to 3D geometry and mathematics. The right panel presents a human reasoning path, demonstrating how visual transformations, rotations, and inside-outside shifts lead to the final answer. We encourage readers to test these examples with MLLM models (e.g., O1 or Gemini) to assess their reasoning capabilities.

II. Metrics

Limitations of Current Methods

The current LLM, when used as a judge, struggles to accurately compare reasoning paths. Even with an accuracy of only 15%, it still assigns high scores. Moreover, even recent reasoning models have difficulty differentiating the O1 response. from the ground truth reasoning (GTR).

GTR represents the ground truth reasoning path, while GTR+ is a refined, structured extension of GTP and GTP+P that incorporates human-labeled visual pattern annotations.

Reasoning Path

We divide the reasoning process into four key stages inspired by human visual reasoning: perception, recognition, abstraction, and deduction. Unlike general visual tasks, where perception involves detecting raw visual features, humans often have implicit perception because the provided visual elements are already structured for direct recognition of useful components. Even for shown complex problems, a model with strong visual abilities—like Gemini—can effectively analyze patterns and logical structures to determine the correct answer.

Reasoning Fidelity

This diagram illustrates three proposed evaluation metrics: Agreement, which extracts common elements across model answers and compares their consistency; Match, which assesses alignment between model responses and ground truth across Recognition, Abstraction, and Deduction; and Perception Similarity, which measures the similarity of extracted perception attributes across different models.

III. Result

The first row visualizes the correlation between Recognition (R) and Abstraction (A), while the second examines the transition from Abstraction (A) to Deduction (D). Bubble sizes represent confidence levels. This figure highlights inter-model variability, with stronger correlations in R → A than A → D, indicating challenges in applying logical deduction even when abstraction is well-formed.

Model Performance and Alignment

We have also included a radar chart for a quick comparison of model performance, highlighting variations across key reasoning and perception dimensions. Notably, models such as Qwen2.5 and GPT-4o show higher agreement scores, suggesting a stronger consensus in visual perception, while o1 achieves relatively higher abstraction and deduction scores, demonstrating superior reasoning capabilities.
On the right, it's a pairwise similarity matrix comparing model responses. Notably, gemini and 4o exhibit the strongest perceptual alignment, suggesting these models process visual data in a similar way. In contrast, qvq shows the weakest alignment with both llava and O1, indicating a distinctly different interpretation of visual inputs. Additionally, 4o and qwen display relatively strong alignment with several models, highlighting their broader generalizability, while O1 maintains moderate alignment without dominating similarity with any one model.

IV. More Qualitative results

This is a detailed case study that explores the question from multiple perspectives, including reasoning paths from various models and insights from human analysis. We have identified a few recurring mistake patterns:

  1. Misinterpreting Features: Errors stem from inaccurate element counts or misestimating numerical patterns.
  2. Reliance on Intuition: Mistakes occur when decisions are based on surface-level visual trends rather than logical transformations.
  3. Failure to Verify All Options: Incorrect answers can result from choosing seemingly correct options without considering all possibilities.