Vision-language models (VLMs) have made rapid progress in multimodal reasoning, but they still fall short when tasks require the model to notice small or easily overlooked regions in an image. That limitation is the motivation for a new preprint on arXiv titled 'Look Before You Reason: Implicit Visual Thinking for Efficient Multimodal Reasoning.'
The authors argue that recent 'thinking with images' approaches, which typically involve generating intermediate image representations, are one direction, but they propose an alternative: implicit visual thinking. Rather than explicit intermediate steps, the model conducts visual reasoning internally, which the authors suggest is more efficient.
The abstract positions this as a way to improve fine-grained visual understanding while keeping multimodal reasoning efficient. However, the announcement gives no quantitative results, so the actual gains are not yet known.