Causal-LLaVA: Causal Disentanglement for Mitigating Hallucination in Multimodal Large Language Models

TL;DR

Causal-LLaVA introduces a causality-based framework with intervention modules to reduce object hallucinations in multimodal large language models, improving factual accuracy without sacrificing performance.

Contribution

It presents a novel causality-driven disentanglement approach with specific modules to mitigate hallucinations caused by dataset biases in MLLMs.

Findings

Significantly reduces object hallucinations in MLLMs.

Maintains high performance on multimodal benchmarks.

Visualization confirms better object representation separability.

Abstract

Multimodal Large Language Models (MLLMs) have demonstrated strong performance in visual understanding tasks, yet they often suffer from object hallucinations--generating descriptions of objects that are inconsistent with or entirely absent from the input. This issue is closely related to dataset biases, where frequent co-occurrences of objects lead to entangled semantic representations across modalities. As a result, models may erroneously activate object representations that are commonly associated with the input but not actually present. To address this, we propose a causality-driven disentanglement framework that mitigates hallucinations through causal intervention. Our approach includes a Causal-Driven Projector in the visual pathway and a Causal Intervention Module integrated into the final transformer layer of the language model. These components work together to reduce spurious correlations caused by biased training data. Experimental results show that our method significantly reduces hallucinations while maintaining strong performance on multiple multimodal benchmarks. Visualization analyses further confirm improved separability of object representations. The code is available at: https://github.com/IgniSavium/Causal-LLaVA