ReLoop: “Seeing Twice and Thinking Backwards” via Closed-loop Training to Mitigate Hallucinations in Multimodal understanding

Jianjiang Yang, Yanshu Li, Ziyan HuangEMNLP 2025

TL;DR

This paper introduces ReLoop, a novel closed-loop training framework for multimodal large language models (MLLMs) that significantly reduces hallucinations by enabling models to "see twice and think backwards." By integrating semantic reconstruction, visual description, and attention alignment, ReLoop encourages internal consistency, leading to more reliable and factually grounded outputs.

What Problem Does This Paper Address?

Multimodal Large Language Models (MLLMs) have made impressive strides in tasks like visual question answering (VQA). However, they often suffer from "hallucinations" – generating outputs that contradict or misrepresent the input image or question. This is a critical issue for the reliability of MLLMs in real-world applications. Existing methods often rely on external checks after generation, lacking an internal mechanism to prevent these errors during training. This paper addresses this gap by proposing a training framework that imbues MLLMs with the ability to self-validate their outputs.

Methods

ReLoop employs a "closed-loop" training process inspired by human cognitive processes. After an MLLM generates an initial answer (the "first look"), ReLoop initiates a "reflection" phase. This involves:

  • Consistency Feedback Plugin (CFP): This plugin, composed of frozen pre-trained modules, performs two key tasks:
  • CFP-Lang (Semantic Reconstruction): Reconstructs a proxy question from the answer and image to check if the model understood the original question.
  • CFP-Vis (Visual Description): Generates a description of the image based on the answer to verify visual grounding.
  • Attention Supervision: This module analyzes the model's internal attention maps to ensure it's focusing on the relevant parts of the image.

These feedback signals are aggregated into a loss function that guides the model to "see twice and think backwards" – first to answer, then to reassess and correct its output. An Adaptive Consistency Weighting (ACW) mechanism is also introduced to dynamically adjust the influence of feedback signals, preventing noisy or nonsensical answers from negatively impacting training.

Datasets

The paper uses the LLaVA-Instruct-150K dataset for training. For evaluation, they employ a curated set of 500 QA-image pairs from the filtered validation set, split equally between hallucinated and non-hallucinated cases. They also evaluate on several established benchmarks: POPE, CHAIR, AMBER, MMHal-B, HallusionBench, MME, Faith, and FaithS.

Key Insights and Ideas

  • Human-like "Reversible Thinking": The core idea of emulating human cognitive processes, where we reassess our answers, is a powerful way to tackle hallucinations.
  • Closed-Loop Training for Internal Consistency: Instead of external checks, ReLoop builds self-correction directly into the training loop, making it more efficient and robust.
  • Frozen Feedback Modules as Supervisors: Repurposing pre-trained models as frozen "supervisors" (CFPs) is an effective way to leverage existing capabilities without the risk of catastrophic forgetting or overfitting.
  • Multi-Signal Supervision is Key: Combining semantic, visual, and attention consistency signals provides a more comprehensive check than any single signal alone.
  • Adaptive Weighting for Robustness: The ACW mechanism is crucial for handling imperfect or noisy feedback, making the training process more stable.

Key Results

ReLoop significantly reduces hallucination rates across various benchmarks and hallucination types (Object, Attribute, Relation, Event). For instance, on MiniGPT-4, ReLoop reduced hallucination rates by 14.2% overall. It also enhances cross-modal faithfulness, with F1 scores increasing by up to 10.6% on MiniGPT-4. The method demonstrates robustness to noisy supervision and nonsensical initial answers. Ablation studies confirm the importance of each component, with the full ReLoop framework achieving the best performance.

Limitations and Future Work

The authors acknowledge that ReLoop's effectiveness on relation and event hallucinations is more modest compared to object and attribute hallucinations, suggesting that these types might require higher-order reasoning or specialized supervision. They also note that ReLoop relies on paired image-question-answer data, which could be a challenge in low-resource domains. Future work could explore reducing this supervision dependency, adapting ReLoop to domains like dense OCR or charts, and investigating its performance on temporal reasoning tasks.

Summary updated

BibTeX
@inproceedings{yang2025reloop,
  author = {Jianjiang Yang and Yanshu Li and Ziyan Huang},
  title = {{ReLoop: “Seeing Twice and Thinking Backwards” via Closed-loop Training to Mitigate Hallucinations in Multimodal understanding}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure