Logical Closed Loop: Uncovering Object Hallucinations in Large Vision-Language Models

TL;DR

Object hallucination in LVLMs poses reliability and safety challenges. The paper introduces LogicCheckGPT, a training-free, plug-and-play system that detects and mitigates hallucinations by probing logical consistency through an object–attribute–object closed loop, quantified by the loop rate $\mathcal{S}(o_i) = \frac{1}{n_i}\sum_{j=1}^{n_i} x_{i,j}$ with threshold $\lambda$. The method comprises five stages—object extraction, object-to-attribute inquiring, attribute-to-object inquiring, loop checking, and mitigation—applied across multiple LVLMs and benchmarks, achieving substantial improvements over baselines. Results on POPE, MME, and GPT-4v-assisted evaluation demonstrate robustness and generality, underscoring the practical impact of leveraging intrinsic model reasoning to curb object hallucinations without additional training. The work offers interpretable, deployable strategies for safer LVLM deployments and lays groundwork for extending logical-consistency checks to broader hallucination types.

Abstract

Object hallucination has been an Achilles' heel which hinders the broader applications of large vision-language models (LVLMs). Object hallucination refers to the phenomenon that the LVLMs claim non-existent objects in the image. To mitigate the object hallucinations, instruction tuning and external model-based detection methods have been proposed, which either require large-scare computational resources or depend on the detection result of external models. However, there remains an under-explored field to utilize the LVLM itself to alleviate object hallucinations. In this work, we adopt the intuition that the LVLM tends to respond logically consistently for existent objects but inconsistently for hallucinated objects. Therefore, we propose a Logical Closed Loop-based framework for Object Hallucination Detection and Mitigation, namely LogicCheckGPT. In specific, we devise logical consistency probing to raise questions with logical correlations, inquiring about attributes from objects and vice versa. Whether their responses can form a logical closed loop serves as an indicator of object hallucination. As a plug-and-play method, it can be seamlessly applied to all existing LVLMs. Comprehensive experiments conducted on three benchmarks across four LVLMs have demonstrated significant improvements brought by our method, indicating its effectiveness and generality.

Figures (15)

• Figure 1: An example of object hallucinations. Hallucinated objects are highlighted in red. The LVLM shows different logical consistency to hallucinated object "apple" and existent object "banana".
• Figure 2: The proposed framework LogicCheckGPT. For LVLM responses to multimodal instructions, LogicCheckGPT employs the following five steps to alleviate object hallucinations: object extraction, object-to-attribute inquiring, attribute-to-object inquiring, logical close loop checking, and hallucination detection and mitigation.
• Figure 3: The visualization of two representative examples of our LogicCheckGPT for mPLUG-Owl. The hallucinated objects are highlighted in red and attributes are highlighted in magenta.
• Figure 4: The performance comparison between LogicCheckGPT and several variants, vanilla, FreeCheck, LogicCheckGPT w/o AOP (w/o AOP) and LogicCheckGPT w/o LCL (w/o LCL) across LVLMs on POPE adversarial setting.
• Figure 5: The performance of different threshold $\lambda$.