LENS: Learning to Segment Anything with Unified Reinforced Reasoning
Proceedings of the AAAI Conference on Artificial Intelligence · 14.03.2026 · 40 (16) · 13952-13960
DOI: 10.1609/aaai.v40i16.38405 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence, online erschienen am 14.03.2026. Verfasst von Lianghui Zhu und 10 weiteren Autor:innen. Die Arbeit wurde bisher 3-mal zitiert (Crossref).
Abstract (Originalsprache)
Text-prompted image segmentation enables fine-grained visual understanding and is critical for applications such as human-computer interaction and robotics. However, existing supervised fine-tuning methods typically ignore explicit chain-of-thought (CoT) reasoning at test time, which limits their ability to generalize to unseen prompts and domains. To address this issue, we introduce LENS, a scalable reinforcement-learning framework that jointly optimizes the reasoning process and segmentation in an end-to-end manner. We propose unified reinforcement-learning rewards that span sentence-, box-, and segment-level cues, encouraging the model to generate informative CoT rationales while refining mask quality. Using a publicly available 3-billion-parameter vision–language model, i.e., Qwen2.5-VL-3B-Instruct, LENS achieves an average cIoU of 81.2% on the RefCOCO, RefCOCO+, and RefCOCOg benchmarks, outperforming the strong fine-tuned method, i.e., GLaMM, by up to 5.6%. These results demonstrate that RL-driven CoT reasoning significantly enhances text-prompted segmentation and offers a practical path toward more generalizable Segment Anything models (SAM).
Zitieren
APA
Zhu, L., Ouyang, B., Zhang, Y., Cheng, T., Hu, R., Shen, H., Ran, L., Chen, X., Yu, L., Liu, W., & Wang, X. (2026). LENS: Learning to Segment Anything with Unified Reinforced Reasoning. Proceedings of the AAAI Conference on Artificial Intelligence, 40(16), 13952-13960. https://doi.org/10.1609/aaai.v40i16.38405
Vancouver
Zhu L, Ouyang B, Zhang Y, Cheng T, Hu R, Shen H, et al. LENS: Learning to Segment Anything with Unified Reinforced Reasoning. Proceedings of the AAAI Conference on Artificial Intelligence. 2026;40(16):13952-13960. doi:10.1609/aaai.v40i16.38405
BibTeX
@article{Zhu2026,
author = {Zhu, Lianghui and Ouyang, Bin and Zhang, Yuxuan and Cheng, Tianheng and Hu, Rui and Shen, Haocheng and Ran, Longjin and Chen, Xiaoxin and Yu, Li and Liu, Wenyu and Wang, Xinggang},
title = {LENS: Learning to Segment Anything with Unified Reinforced Reasoning},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
year = {2026},
volume = {40},
number = {16},
pages = {13952--13960},
doi = {10.1609/aaai.v40i16.38405}
}
EUVIMED