The paper introduces PIVOT, a dual-level learning framework designed to improve visually-grounded multimodal reasoning in large vision-language models. PIVOT employs a self‑calibrated experience replay mechanism to selectively reuse valuable visual reasoning trajectories, and a vision‑guided advantage allocation scheme that assigns extra rewards to tokens with strong visual support. Experiments on multiple benchmarks show that PIVOT enhances the multimodal reasoning performance of these models.
By Xinxin Song, Siyuan Li, Tingxiong Xiao, Jinli Suo
The paper introduces Activation Replay, a training‑free method that improves reasoning in post‑trained large multimodal models (LMMs) by replaying low‑entropy activations from the base model’s input context. It shows that Reinforcement Learning with Verifiable Rewards (RLVR) shifts low‑entropy activations and that modulating these activations enhances reasoning across tasks such as mathematics, visual agents, and video reasoning. Experiments demonstrate that Activation Replay outperforms alternatives like high‑entropy replay or direct cross‑model intervention, boosting Pass@K and broadening RLVR’s reasoning coverage.
By Yun Xing, Xiaobin Hu, Qingdong He, Jiangning Zhang, Shuicheng Yan, Shijian Lu, Yu-Gang Jiang
PEAR is an adaptive advantage reshaping method designed to improve perception and exploration in large vision‑language models. It builds on visual dependency as a base signal and adds an entropy‑based exploration gain only at high‑entropy positions, giving larger gains where visual dependency is lower. Experiments on Qwen2.5‑VL‑3B/7B demonstrate that PEAR outperforms baseline methods across multimodal reasoning benchmarks and achieves leading results on textual reasoning and general multimodal tasks.
By Junxin Wang, Yingfeng Luo, Xiangchen Song, Weiqiao Shan, Runsong Zhao, Xinyu Liu, Shunjie Xing, Yifu Huo, Dingyang Lin, Qiuyu Gu, Yongyu Mu, Chenglong Wang, Changliang Li, Jingbo Zhu, Tong Xiao
The paper introduces Perception and Reasoning Attribution (PARA), a method that quantifies how much of a vision‑language model’s errors can be fixed by better perception versus stronger reasoning. PARA reveals that reinforcement learning (RL) benefits more from improving reasoning than perception, and motivates a new group‑relative attention distillation objective called GRAD. GRAD directly guides weaker rollouts toward the visual attention patterns of the best rollouts, consistently boosting performance across multiple model scales and benchmarks.
By Zeyuan Yang, Zhuohao Chen, Jiaben Chen, Yao Zhang, Maohao Shen, Chuang Gan
arXiv:2607. 14682v1 Announce Type: new Abstract: Efficient multimodal document question answering with explicit visual grounding, locating the precise document region that supports each answer remains an open challenge.
By Harikrishnan P M, Goutham Vignesh, Ganesh Parab, Saisubramaniam Gopalakrishnan, Vishal Vaddina, Varun V, Rohit Agrawal
The paper introduces VIG (Visual Information Gain), an information‑theoretic reward that evaluates each token in a multimodal chain‑of‑thought by measuring how much the image reduces its predictive uncertainty. VIG is computed online using two forward passes—one with and one without the image—eliminating the need for reference chains or external annotations. Experiments on six multimodal reasoning benchmarks and multiple Qwen3‑VL‑Thinking model sizes show that VIG consistently improves the accuracy–efficiency trade‑off, demonstrating that efficient multimodal reasoning arises from increasing visual information density rather than merely limiting chain length.
By Wen Luo, Xiaohan Yi, Xiaotao Huang, Liqun Huang
The paper introduces GRIT, a training framework that adds visual grounding to reinforcement learning with verifiable rewards (RLVR) for vision-language models, addressing the issue of Correct-but-Visually-Misgrounded Reasoning (CVMR). CVMR occurs when models produce correct final answers while misperceiving visual information during intermediate reasoning steps, a problem that standard RLVR rewards cannot detect. By analyzing token-level attention patterns, the authors show that CVMR tokens differ from well-grounded tokens and use this insight to guide GRIT, which improves both reasoning accuracy and grounding fidelity. Additionally, the authors present GRIT-Bench, a new benchmark for evaluating text-rich reasoning across varying text lengths and reasoning depths.
By Chanseul Cho, Noori Bae, Dongheon Lee, Youngmin Ro
VideoLVR introduces a method for multimodal large language models to maintain visual grounding during chain-of-thought reasoning by constructing latent visual evidence from local video regions at relevant moments. It uses a visual token prediction objective that unifies visual evidence reconstruction and text generation, allowing models to interleave text tokens with latent visual states. The approach is further extended with reinforcement learning, treating visual tokens as actions and optimizing policies with final-answer rewards, thereby improving video understanding without expanding the multimodal context.
By Shilong Li, Xingyuan Bu, Hao Shi, Yun Xing, Weixin Feng, Qiangpeng Yang, Shilei Wen
arXiv:2609.21675v1 Announce Type: new
Abstract: Despite the remarkable progress in Multimodal Large Language Models (MLLMs), prevailing Chain-of-Thought (CoT) paradigms remain confined to the natural...
By Wan Xu, Yuanfan Guo, Kevin Han, LaLa Chen, Wangmeng Zuo
arXiv:2602.08241v2 Announce Type: replace
Abstract: While chain-of-thought (CoT) reasoning has substantially improved multimodal large language models (MLLMs) on complex reasoning tasks, existing app...
By Siqu Ou, Tianrui Wan, Zhiyuan Zhao, Junyu Gao, Xuelong Li
The paper investigates the faithfulness of visual information in multimodal chain-of-thought (MCoT) reasoning by large vision-language models (LVLMs). It shows that visual cues in MCoT are often inaccurate and largely ignored, as predictions change little when visual input is altered but significantly when textual input is changed. To improve visual faithfulness, the authors propose an annotation‑free Sufficient‑Component Cause Model (SCCM) that encourages the generation of minimal yet sufficient visual components, and demonstrate its effectiveness across fine‑grained perception and reasoning benchmarks.
By Zujing Liu, Junwen Pan, Qi She, Yuan Gao, Guisong Xia
The paper investigates how reinforcement learning with verifiable rewards can improve multimodal large language models (LLMs) by separating perception and reasoning stages. Experiments on algorithmic visual puzzles show that perception, not reasoning, limits performance, and replacing images with textual descriptions boosts Claude models by over 20 points. Reward designs that encourage visual grounding during reasoning, such as GRPO for Qwen-2.5-VL-7B, yield structured reasoning and a 5.56‑point gain, though benefits vary across categories and trade‑offs exist between cross‑domain transfer and in‑domain accuracy.
By Omar Sharif, Eftekhar Hossain, Nikhil Singh, Patrick Ng