Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Piculet: Specialized Models-Guided Hallucination Decrease for MultiModal Large Language Models

Kohou Wang, Xiang Liu, Zhaoxiang Liu, Kai Wang, Shiguo Lian

TL;DR

The paper tackles hallucinations in Multimodal Large Language Models by introducing Piculet, a training-free preprocessing framework that uses specialized lightweight models to extract factual image details (objects, text, faces) and feeds them as structured external knowledge alongside the image and user query to the MLLM. This approach avoids retraining and large LLM dependencies, requiring only a single MLLM inference plus small models, making it efficient and broadly compatible. Experimental results on POPE, MMEMME, and LLaVA-QA90 show consistent improvements across object existence, counts, OCR, color, and celebrity recognition tasks, outperforming both training-based methods and existing training-free baselines like Woodpecker. Overall, Piculet offers a practical, scalable solution to mitigate visual hallucination in MLLMs with potential for widespread deployment in real-world multimodal AI systems.

Abstract

Multimodal Large Language Models (MLLMs) have made significant progress in bridging the gap between visual and language modalities. However, hallucinations in MLLMs, where the generated text does not align with image content, continue to be a major challenge. Existing methods for addressing hallucinations often rely on instruction-tuning, which requires retraining the model with specific data, which increases the cost of utilizing MLLMs further. In this paper, we introduce a novel training-free method, named Piculet, for enhancing the input representation of MLLMs. Piculet leverages multiple specialized models to extract descriptions of visual information from the input image and combine these descriptions with the original image and query as input to the MLLM. We evaluate our method both quantitively and qualitatively, and the results demonstrate that Piculet greatly decreases hallucinations of MLLMs. Our method can be easily extended to different MLLMs while being universal.

Piculet: Specialized Models-Guided Hallucination Decrease for MultiModal Large Language Models

TL;DR

The paper tackles hallucinations in Multimodal Large Language Models by introducing Piculet, a training-free preprocessing framework that uses specialized lightweight models to extract factual image details (objects, text, faces) and feeds them as structured external knowledge alongside the image and user query to the MLLM. This approach avoids retraining and large LLM dependencies, requiring only a single MLLM inference plus small models, making it efficient and broadly compatible. Experimental results on POPE, MMEMME, and LLaVA-QA90 show consistent improvements across object existence, counts, OCR, color, and celebrity recognition tasks, outperforming both training-based methods and existing training-free baselines like Woodpecker. Overall, Piculet offers a practical, scalable solution to mitigate visual hallucination in MLLMs with potential for widespread deployment in real-world multimodal AI systems.

Abstract

Multimodal Large Language Models (MLLMs) have made significant progress in bridging the gap between visual and language modalities. However, hallucinations in MLLMs, where the generated text does not align with image content, continue to be a major challenge. Existing methods for addressing hallucinations often rely on instruction-tuning, which requires retraining the model with specific data, which increases the cost of utilizing MLLMs further. In this paper, we introduce a novel training-free method, named Piculet, for enhancing the input representation of MLLMs. Piculet leverages multiple specialized models to extract descriptions of visual information from the input image and combine these descriptions with the original image and query as input to the MLLM. We evaluate our method both quantitively and qualitatively, and the results demonstrate that Piculet greatly decreases hallucinations of MLLMs. Our method can be easily extended to different MLLMs while being universal.
Paper Structure (13 sections, 5 figures, 4 tables)

This paper contains 13 sections, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. MLLMs' Hallucinations
  4. Training-based Methods
  5. Training-free Methods
  6. Method
  7. Specialized Models
  8. Input Formulation
  9. Experiments
  10. Datasets
  11. Experimental Results
  12. Ablation Study.
  13. Conclusion

Figures (5)

  • Figure 1: Illustration of Hallucination of MLLMs. This MLLM generates descriptions of an image with wrong information, including the number of people and cups, and also, there is only whiteboard, not a blackboard on the left wall.
  • Figure 2: Details of question formulation. In each sub-image, we adopt a randomly chosen image to exemplify the concrete operation.
  • Figure 3: Illustration of our method's processing. Red words are pre-processed results of specialized models, yellow words are the predefined prompt everybody usually uses, and blue words are the user's original query, purple words are model's reply without hallucination. The recognized characters, faces and objects are integrated into one single sentence, which, alongside the user's original query and image, serves as the final input of MLLMs.
  • Figure 4: Flowchart of our method. Given an image and a query, firstly we utilize specialized models to extract descriptions of visual information, these descriptions are then reorganized by prompt formulation block and combined with the original user's query, the newly combined query and image are then input into the MLLM.
  • Figure 5: Prompt template for GPT-4V-aided evaluation. Response 1 and Response 2 are the original responses and the corrected ones, respectively.