How Far Are We from Intelligent Visual Deductive Reasoning?

Even state-of-the-art Vision-Language Models (VLMs) like GPT-4V face hurdles in visual deductive reasoning, a vital cognitive capability. This work delves into their performance on complex tasks such as Raven’s Progressive Matrices, revealing a significant divide between text and image reasoning abilities.

• Highlights inadequacies of VLMs in pattern recognition for RPM tasks.
• Questions the direct applicability of successful text reasoning strategies.
• Urges for further advancements in multi-hop relational and deductive visual reasoning.

These findings indicate that while LLMs show promise in text-based reasoning, our journey towards visual reasoning intelligence is still nascent. Fostering VLMs that can effectively navigate abstract visual cues is quintessential for the next wave of AI breakthroughs.