Mitigating Multimodal Hallucination via Phase-wise Self-reward

TL;DR

This paper introduces PSRD, a dynamic inference-time framework that mitigates vision hallucination in LVLMs by using phase-wise self-reward signals and a lightweight reward model, significantly reducing hallucinations.

Contribution

It proposes a novel phase-wise self-reward decoding method that dynamically suppresses hallucinations without external supervision, improving efficiency and effectiveness.

Findings

PSRD reduces hallucination rate of LLaVA-1.5-7B by 50%.

It outperforms existing post-hoc methods across five benchmarks.

The approach enables a controllable trade-off between performance and inference efficiency.

Abstract

Large Vision-Language Models (LVLMs) still struggle with vision hallucination, where generated responses are inconsistent with the visual input. Existing methods either rely on large-scale annotated data for fine-tuning, which incurs massive computational overhead, or employ static post-hoc strategies that overlook the dynamic nature of hallucination emergence. To address these, we introduce a new self-rewarding framework, enabling dynamic hallucination mitigation at inference time without external supervision. On the empirical side, we reveal that visual hallucination exhibits phase-wise dynamic patterns, peaking at the onset of each semantic phase. Drawing on these insights, we propose \textbf{PSRD} (\textbf{Phase-wise \textbf{S}elf-\textbf{R}eward \textbf{D}ecoding) for online hallucination correction guided by phase-wise self-reward signals. To reduce the cost of repeated self-evaluation during decoding, we distill the hallucination guidance signal from LVLMs into a lightweight reward model. The reward model subsequently provides on-the-fly guidance for targeted intervention during the decoding process, enabling precise hallucination suppression. The proposed PSRD significantly reduces the hallucination rate of LLaVA-1.5-7B by 50.0% and consistently outperforms existing post-hoc methods across five hallucination evaluation benchmarks for four LVLMs. Further analysis confirms that PSRD effectively mitigates hallucination propagation and achieves a highly controllable trade-off between strong performance and inference efficiency.