Online Self-Calibration Against Hallucination in Vision-Language Models

TL;DR

This paper introduces OSCAR, an online self-calibration framework for vision-language models that reduces hallucinations by leveraging a generative-discriminative gap and iterative refinement.

Contribution

It proposes a novel online self-calibration method using Monte Carlo Tree Search and a dual-granularity reward to improve hallucination mitigation in LVLMs.

Findings

OSCAR achieves state-of-the-art hallucination reduction on benchmarks.

The method improves general multimodal capabilities.

It effectively aligns model perception with visual input.

Abstract

Large Vision-Language Models (LVLMs) often suffer from hallucinations, generating descriptions that include visual details absent from the input image. Recent preference alignment methods typically rely on supervision distilled from stronger models such as GPT. However, this offline paradigm introduces a Supervision-Perception Mismatch: the student model is forced to align with fine-grained details beyond its perceptual capacity, learning to guess rather than to see. To obtain reliable self-supervision for online learning, we identify a Generative-Discriminative Gap within LVLMs, where models exhibit higher accuracy on discriminative verification than open-ended generation. Leveraging this capability, we propose \textbf{O}nline \textbf{S}elf-\textbf{CA}lib\textbf{R}ation (OSCAR), a framework that integrates Monte Carlo Tree Search with a Dual-Granularity Reward Mechanism to construct preference data and iteratively refines the model via Direct Preference Optimization. Extensive experiments demonstrate that OSCAR achieves state-of-the-art performance on hallucination benchmarks while improving general multimodal capabilities.