Decoding by Perturbation: Mitigating MLLM Hallucinations via Dynamic Textual Perturbation

arXiv cs.CL / 4/15/2026

💬 OpinionSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper argues that multimodal LLM hallucinations often arise because language priors over-dominate visual evidence during decoding, especially in how visual grounding reacts to textual phrasing.
It proposes “Decoding by Perturbation (DeP),” a training-free method that introduces controlled, multi-level textual perturbations during inference to elicit and manage latent language priors.
DeP uses attention variance to reinforce stable, evidence-aligned regions in the feature space while suppressing suspicious noise.
It further estimates an interpretable “prior drift direction” from logits statistics to counteract probability biases caused by textual co-occurrences.
Experiments reportedly show DeP reduces hallucinations and improves benchmark performance across multiple evaluations.

Abstract

Multimodal Large Language Models frequently suffer from inference hallucinations, partially stemming from language priors dominating visual evidence. Existing training-free mitigation methods either perturb the visual representation and deviate from the natural image distribution, or enforce intrusive manipulations that compromise the model's inherent generative fluency. We introduce a novel perspective that multimodal hallucination manifests as the hypersensitivity of visual grounding to textual phrasing during the decoding phase. Building on this insight, we propose Decoding by Perturbation (DeP), a training-free framework mitigating prior-induced hallucinations via controlled textual interventions. DeP employs a dynamic probe applying multi-level textual perturbations to elicit latent language priors. Leveraging attention variance, it enhances stable evidence regions while suppressing suspicious noise in the feature space. Furthermore, it constructs an interpretable prior drift direction using logits statistics to counteract probability biases from textual co-occurrences. Extensive experiments confirm DeP effectively reduces hallucinations and achieves superior performance across multiple benchmarks.