Back to feed

Vision-driven Preference Synthesis for Mitigating Hallucinations in VLMs

Jun 2026 · arXiv.org · Vol abs/2606.28401 · 0 citations · 96 references
Computer Science

TL;DR

ViPSy (Vision-driven Preference Synthesis), a framework for constructing preference data that are both policy-aligned and visually grounded, is proposed, and experiments show that the resulting VLM, preference-aligned with ViPSy-constructed preference pairs, achieves a new state-of-the-art in hallucination mitigation.

Abstract

Vision-Language Models (VLMs) have shown strong performance in visual understanding, yet they still suffer from hallucinations, generating content that is not grounded in the image. Preference alignment is a promising approach to improve visual faithfulness, but its success depends heavily on how preference pairs are constructed. Existing methods exhibit two key limitations; (a) intervention-based methods often introduce significant deviation from the policy distribution, and (b) sampling-based methods often underuse visual information during the construction. In this paper, we propose ViPSy (Vision-driven Preference Synthesis), a framework for constructing preference data that are both policy-aligned and visually grounded. Our framework consists of two stages; in the first stage, ViPSy derives a visual cue from recurring object-level content across semantically aligned image variants, so preference construction can rely on visual information rather than language priors. In the second stage, ViPSy conditions the policy's own rollouts on this cue, allowing candidates to be guided by visually grounded content while staying close to the policy's response distribution. The resulting candidates remain close to the policy's response distribution while better leveraging visual information from the image. Experiments show that the resulting VLM, preference-aligned with ViPSy-constructed preference pairs, achieves a new state-of-the-art in hallucination mitigation. Compared with the previous state-of-the-art method, it reduces hallucination rates on AMBER and Object HalBench by 35.7% and 24.5%, respectively. The resulting model further improves on general visual grounding benchmarks, e.g., MMStar, MMVP, and CV-Bench, while also yielding gains in semantic segmentation and ImageNet linear probing, underscoring the effectiveness of our framework in enhancing the model's visual capabilities.

View source

Similar papers

Preprint Jul 2026

Role-Break in Attention Heads: Understanding and Detecting Hallucinations in VLMs

A lightweight linear detector is built on top of Role-Break that requires no fine-tuning of the VLM, whose feature dimension stays below 5,000 and reaches an average AUROC of 93.23 across six VLMs and four benchmarks.

Mingyu Wang, Weilin Jin, Wenbo Li et al. · 0 citations
Jun 2026

Clearer Sight, Fewer Lies: Oriented Pickup Preference Optimization for Multimodal Hallucination Mitigation

Multimodal Large Language Models (MLLMs) are prone to hallucination as their generation preferences are insufficiently calibrated to visual evidence, causing them to fall back on linguistic priors, rather than faithful grounding. In this work, we start from an empirical observation: when query-relevant visual evidence is explicitly strengthened using the model's own attention, generation becomes more accurate, suggesting that many failures do not arise solely from missing perception, but from an insufficient tendency to trust the evidence the model has already attended to. Motivated by this finding, we propose Oriented Pickup Preference Optimization (\texttt{OPPO}), an evidence-aware alignment objective that learns preferences over the strength of visual evidence, rather than only response quality. Concretely, \texttt{OPPO} contrasts the same faithful response under stronger, anchored, weaker-evidence views, turning naive visual preference into ordered visual-evidence alignment. We further combine this objective with fine-grained span-level and token-level regularization to stabilize the training. Besides, we provide a theoretical analysis showing that ordered evidence margins induce a positive lower bound on local visual sensitivity. Extensive evaluations across hallucination and general-purpose benchmarks demonstrate that \texttt{OPPO} consistently outperforms baseline methods.

Xin Zou, Hao Deng, Yibo Yan et al. · 0 citations

RIVS: Mitigating Hallucination in Large Vision-Language Models via Representation Intervention on Visual Grounding Shift

This work studies hallucination from the perspective of dynamic representation shift during generation and proposes an online projection-based intervention on intermediate hidden states to suppress the hallucination-related directions, mitigating hallucinations while preserving language quality.

Xuanyu Yin, Xiaoye Qu, ∗. WeiWei · 0 citations
Preprint Aug 2026

Test-Time Hallucination Control in Large Vision-Language Models

Object Hallucination in large vision-language models (LVLMs), where models generate non-factual content about input images, remains a critical barrier to their reliability in real-world applications. Existing mitigation strategies can be categorized into training-based and training-free methods. Training-based methods often achieve strong performance but are costly, requiring extensive computational resources, large-scale data, and time-consuming fine-tuning. Training-free approaches are particularly appealing due to their efficiency. However, existing training-free methods either require multiple decoding rounds, which adds computational overhead, or modify internal states in a model-specific way that risks degrading pretrained knowledge. We propose Test-Time Hallucination Mitigation (TTH) method, a novel training-free method that addresses both limitations. TTH introduces a token-validator module, implemented as a zero-shot Multi-Modal Classifier (MMC), to generate auxiliary logits grounded in the input image. These logits are fused with the original LVLM outputs at the token level for object tokens selected from a candidate pool. An entropy-based weighting scheme is then applied to enable robust and accurate predictions. Extensive experiments across multiple LVLM families and diverse benchmarks demonstrate that TTH consistently improves accuracy and robustness, underscoring its generalizability and practical effectiveness. Code is released at https://github.com/Mehran-TAM/TTH

Mehran Tamjidi, Hamidreza Dastmalchi, Ali Cheraghian et al. · 0 citations
Open access 2026

DIVE: A Training-free Hallucination Mitigation Mechanism for Complex Scenes

: When facing real-world scenes that are densely populated with objects or contain complex occlusions, Vision-Language Models are often constrained by the language prior in autoregressive decoding, producing severe hallucination phenomena. To address this pain point that limits the reliable deployment of large models, this paper proposes Dual-branch Inference for Visual-prior Elimination, a training-free hallucination mitigation mechanism for complex scenes. By constructing a dual-branch inference structure at the inference stage and introducing a dynamic visual-confidence penalty, this mechanism effectively quantifies and suppresses the overconfidence in the content generation process, forcing the model’s output to be deeply aligned with the underlying visual features. Results on the object hallucination evaluation benchmark POPE show that, without consuming computing power for model fine-tuning, the proposed method reduces the model’s hallucination rate when objects are dense or complex occlusions exist, and brings a slight improvement in the question-answering accuracy of the model on the MSCOCO and VG datasets.

Shuguo Jiang · 0 citations