CR³: Boosting Compositional Reasoning in MLLMs Through Rule-Based Reinforcement Learning
Proceedings of the AAAI Conference on Artificial Intelligence · 14.03.2026 · 40 (29) · 24927-24935
DOI: 10.1609/aaai.v40i29.39680 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence, online erschienen am 14.03.2026. Verfasst von Shun Qian und 5 weiteren Autor:innen.
Abstract (Originalsprache)
Compositional reasoning is a critical capability for multimodal models, enabling systematic understanding of complex scenes through structured combinations of objects, attributes, and relations. However, existing research on this ability primarily focuses on vision-language models (VLMs, e.g., CLIP and SigLIP), with limited exploration of multimodal large language models (MLLMs). To address this gap, we introduce CR³, a novel framework that enhances compositional reasoning abilities of MLLMs via rule-based reinforcement learning. CR³ leverages rule-based rewards to optimize the MLLM's policy on systematically curated multimodal instruction-following tasks, guided by a model-adaptive dynamic task mixing strategy. Our approach boosts performance by over 19% on three compositional reasoning benchmarks, significantly outperforming supervised fine-tuning (SFT) by at least 12%. Crucially, CR³ demonstrates superior generalization by improving performance on out-of-domain benchmarks where SFT methods degrade, highlighting its effectiveness and data efficiency.
Zitieren
APA
Qian, S., Liu, B., Sun, C., Xie, P., Xu, Z., & Wang, B. (2026). CR³: Boosting Compositional Reasoning in MLLMs Through Rule-Based Reinforcement Learning. Proceedings of the AAAI Conference on Artificial Intelligence, 40(29), 24927-24935. https://doi.org/10.1609/aaai.v40i29.39680
Vancouver
Qian S, Liu B, Sun C, Xie P, Xu Z, Wang B. CR³: Boosting Compositional Reasoning in MLLMs Through Rule-Based Reinforcement Learning. Proceedings of the AAAI Conference on Artificial Intelligence. 2026;40(29):24927-24935. doi:10.1609/aaai.v40i29.39680
BibTeX
@article{Qian2026,
author = {Qian, Shun and Liu, Bingquan and Sun, Chengjie and Xie, Peijin and Xu, Zhen and Wang, Baoxun},
title = {CR³: Boosting Compositional Reasoning in MLLMs Through Rule-Based Reinforcement Learning},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
year = {2026},
volume = {40},
number = {29},
pages = {24927--24935},
doi = {10.1609/aaai.v40i29.39680}
}
EUVIMED