Vision-Language Introspection: Mitigating Overconfident Hallucinations in MLLMs via Interpretable Bi-Causal Steering

TL;DR

This paper introduces Vision-Language Introspection (VLI), a training-free framework that reduces hallucinations in multimodal models by diagnosing and actively correcting visual misinterpretations through interpretable, instance-specific steering.

Contribution

VLI presents a novel, training-free inference method that diagnoses hallucination risks and dynamically corrects visual evidence interpretation in multimodal models.

Findings

Reduces object hallucination rates by 12.67% on MMHal-Bench

Improves accuracy by 5.8% on POPE

Achieves state-of-the-art performance on advanced models

Abstract

Object hallucination critically undermines the reliability of Multimodal Large Language Models, often stemming from a fundamental failure in cognitive introspection, where models blindly trust linguistic priors over specific visual evidence. Existing mitigations remain limited: contrastive decoding approaches operate superficially without rectifying internal semantic misalignments, while current latent steering methods rely on static vectors that lack instance-specific precision. We introduce Vision-Language Introspection (VLI), a training-free inference framework that simulates a metacognitive self-correction process. VLI first performs Attributive Introspection to diagnose hallucination risks via probabilistic conflict detection and localize the causal visual anchors. It then employs Interpretable Bi-Causal Steering to actively modulate the inference process, dynamically isolating visual evidence from background noise while neutralizing blind confidence through adaptive calibration. VLI achieves state-of-the-art performance on advanced models, reducing object hallucination rates by 12.67% on MMHal-Bench and improving accuracy by 5.8% on POPE.