How Far Are We from Intelligent Visual Deductive Reasoning?

TL;DR

This paper evaluates the current state of vision-language models in performing visual deductive reasoning using Raven's Progressive Matrices, revealing significant gaps compared to text-based reasoning and highlighting specific challenges in visual pattern comprehension.

Contribution

It introduces a comprehensive evaluation of VLMs on visual deductive reasoning tasks, exposing blindspots and analyzing the limitations of existing strategies like CoT in visual contexts.

Findings

VLMs underperform on visual deductive reasoning tasks compared to text-based reasoning.

Standard reasoning strategies do not directly transfer effectively to visual reasoning tasks.

VLMs struggle with perceiving and understanding multiple abstract patterns in RPM examples.

Abstract

Vision-Language Models (VLMs) have recently demonstrated incredible strides on diverse vision language tasks. We dig into vision-based deductive reasoning, a more sophisticated but less explored realm, and find previously unexposed blindspots in the current SOTA VLMs. Specifically, we leverage Raven's Progressive Matrices (RPMs), to assess VLMs' abilities to perform multi-hop relational and deductive reasoning relying solely on visual clues. We perform comprehensive evaluations of several popular VLMs employing standard strategies such as in-context learning, self-consistency, and Chain-of-thoughts (CoT) on three diverse datasets, including the Mensa IQ test, IntelligenceTest, and RAVEN. The results reveal that despite the impressive capabilities of LLMs in text-based reasoning, we are still far from achieving comparable proficiency in visual deductive reasoning. We found that certain standard strategies that are effective when applied to LLMs do not seamlessly translate to the challenges presented by visual reasoning tasks. A detailed analysis reveals that VLMs struggle to solve these tasks mainly because they are unable to perceive and comprehend multiple, confounding abstract patterns in RPM examples.