SAVOR is a training framework for multimodal large language models that adds token and answer confidence to the output schema, optimises a Group Relative Policy Optimisation objective to penalise calibration error and poor abstention, and uses the learned confidence at inference to revisit visual evidence only when uncertain. Experiments on POPE, HallusionBench, AMBER, and MMHal-Bench with InternVL3-8B and Qwen3-VL-8B backbones show that SAVOR reduces hallucination while maintaining general capability on MME and MMBench, achieving lower Expected Calibration Error than DPO and decoding baselines.
By Zixiu Ding, Zilin Zhao, Yingjie He, Xinlang Kang, Guansu Wang, Wei Zhang
arXiv:2606. 29984v1 Announce Type: new Abstract: Reinforcement Learning (RL) is an important paradigm for improving the reasoning capabilities of Vision-Language Models (VLMs).
By Peng, Lee, Yin Zhang, Yanglin Zhang, Haonan Wu, Zishan Liu, Ruoxi Zang, Xin Zhu, Jiayin Zheng, Jian Yao, Zefeng Ji, Fei Ma
VEPO (Visual Evidence-guided Preference Optimization) is a training-time method designed to reduce object hallucinations in large vision-language models, especially for low-resource languages. It localizes key visual evidence using English queries and creates preference pairs that differ in evidence sufficiency, encouraging models to favor answers supported by complete visual evidence. Experiments on multilingual datasets (POPE and MME) with Chinese, French, and Hindi demonstrate consistent hallucination reduction and improved perception, with the greatest gains in the lowest-resource language.
By Chengpeng Fu, Xiaocheng Feng, Wenshuai Huo, Yichong Huang, Baohang Li, Duyu Tang, Ting Liu
arXiv:2605. 16411v3 Announce Type: replace-cross Abstract: Hallucination remains a fundamental challenge in vision-language models (VLMs), where autoregressive generation may produce linguistically plausible yet physically inconsistent or visually ungrounded responses due to likelihood maximization under joint probabilistic modeling.
By Qinwu Xu
arXiv:2609.39168v1 Announce Type: new
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has improved the reasoning capabilities of Multimodal Large Language Models (MLLMs), yet existing...
By Zhihan Zhang, Lizi Liao
The paper argues that stochastic decoding, common in large language models, is not ideal for Visual Question Answering (VQA) because VQA is a closed‑ended task with head‑heavy answer distributions and epistemic uncertainty. The authors formalize how model calibration relates to predictive accuracy and identify conditions under which greedy decoding is optimal. Experiments across multiple benchmarks show greedy decoding outperforms stochastic sampling, and a new Greedy Decoding for Reasoning Models further improves multimodal reasoning performance.
By Boqi Chen, Xudong Liu, Yunke Ao, Jianing Qiu