Dual-Uncertainty Guided Policy Learning for Multimodal Reasoning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Liu, Rui, Yu, Dian, Zheng, Tong, Dai, Runpeng, Li, Zongxia, Yu, Wenhao, Liang, Zhenwen, Song, Linfeng, Mi, Haitao, Tokekar, Pratap, Yu, Dong
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908768291258368
author Liu, Rui
Yu, Dian
Zheng, Tong
Dai, Runpeng
Li, Zongxia
Yu, Wenhao
Liang, Zhenwen
Song, Linfeng
Mi, Haitao
Tokekar, Pratap
Yu, Dong
author_facet Liu, Rui
Yu, Dian
Zheng, Tong
Dai, Runpeng
Li, Zongxia
Yu, Wenhao
Liang, Zhenwen
Song, Linfeng
Mi, Haitao
Tokekar, Pratap
Yu, Dong
contents Reinforcement learning with verifiable rewards (RLVR) has advanced reasoning capabilities in multimodal large language models. However, existing methods typically treat visual inputs as deterministic, overlooking the perceptual ambiguity inherent to the visual modality. Consequently, they fail to distinguish whether a model's uncertainty stems from complex reasoning or ambiguous perception, preventing the targeted allocation of exploration or learning signals. To address this gap, we introduce DUPL, a dual-uncertainty guided policy learning approach for multimodal RLVR that quantifies and leverages both perceptual uncertainty (via symmetric KL divergence) and output uncertainty (via policy entropy) to guide policy updates. By establishing an uncertainty-driven feedback loop and employing a dynamic branch prioritization mechanism, DUPL recalibrates the policy advantage to focus learning on states with high perceptual or decisional ambiguity, enabling effective targeted exploration beyond passive data augmentation. Implemented on top of GRPO and evaluated on six multimodal mathematical and general-domain reasoning benchmarks, DUPL improves Qwen2.5-VL 3B and 7B models, achieving accuracy gains of up to 11.2% on visual math tasks and up to 7.1% on general-domain reasoning tasks, while consistently outperforming GRPO. These results demonstrate that dual-uncertainty guided policy learning is an effective and generalizable approach for multimodal RLVR.
format Preprint
id arxiv_https___arxiv_org_abs_2510_01444
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Dual-Uncertainty Guided Policy Learning for Multimodal Reasoning
Liu, Rui
Yu, Dian
Zheng, Tong
Dai, Runpeng
Li, Zongxia
Yu, Wenhao
Liang, Zhenwen
Song, Linfeng
Mi, Haitao
Tokekar, Pratap
Yu, Dong
Artificial Intelligence
Computation and Language
Machine Learning
Reinforcement learning with verifiable rewards (RLVR) has advanced reasoning capabilities in multimodal large language models. However, existing methods typically treat visual inputs as deterministic, overlooking the perceptual ambiguity inherent to the visual modality. Consequently, they fail to distinguish whether a model's uncertainty stems from complex reasoning or ambiguous perception, preventing the targeted allocation of exploration or learning signals. To address this gap, we introduce DUPL, a dual-uncertainty guided policy learning approach for multimodal RLVR that quantifies and leverages both perceptual uncertainty (via symmetric KL divergence) and output uncertainty (via policy entropy) to guide policy updates. By establishing an uncertainty-driven feedback loop and employing a dynamic branch prioritization mechanism, DUPL recalibrates the policy advantage to focus learning on states with high perceptual or decisional ambiguity, enabling effective targeted exploration beyond passive data augmentation. Implemented on top of GRPO and evaluated on six multimodal mathematical and general-domain reasoning benchmarks, DUPL improves Qwen2.5-VL 3B and 7B models, achieving accuracy gains of up to 11.2% on visual math tasks and up to 7.1% on general-domain reasoning tasks, while consistently outperforming GRPO. These results demonstrate that dual-uncertainty guided policy learning is an effective and generalizable approach for multimodal RLVR.
title Dual-Uncertainty Guided Policy Learning for Multimodal Reasoning
topic Artificial Intelligence
Computation and Language
Machine Learning
url https://arxiv.org/abs/2510.01444