Advancing Creative Physical Intelligence in Large Multimodal Models
A new paper introduces MM-CreativityBench, a benchmark designed to evaluate creative problem-solving in large multimodal models. The study reveals that current models often struggle with grounded exploration, leading to errors in identifying relevant entities and parts. The authors propose a method to enhance models' performance by aligning creative tool use with visual evidence.
- ▪MM-CreativityBench is a benchmark for affordance-grounded creative tool use in visually rich environments.
- ▪Current large multimodal models often fail to sustain grounded exploration, resulting in overlooked entities and hallucinated attributes.
- ▪The authors propose affordance-grounded alignment to improve models' preference for attribute-affordance reasoning based on visual evidence.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.26396 (cs) [Submitted on 25 May 2026] Title:Advancing Creative Physical Intelligence in Large Multimodal Models Authors:Cheng Qian, Hyeonjeong Ha, Jiayu Liu, Jeonghwan Kim, Emre Can Acikgoz, Bingxuan Li, Kunlun Zhu, Jiateng Liu, Aditi Tiwari, Zhenhailong Wang, Xiusi Chen, Mahdi Namazifar, Heng Ji View a PDF of the paper titled Advancing Creative Physical Intelligence in Large Multimodal Models, by Cheng Qian and 12 other authors View PDF HTML (experimental) Abstract:Large multimodal models (LMMs) have rapidly advanced in perception and reasoning; however, it remains unclear whether these capabilities generalize to discovering visually grounded solutions in open-ended environments, beyond pattern recognition.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
