CoFi-Dec: Hallucination-Resistant Decoding via Coarse-to-Fine Generative Feedback in Large Vision-Language Models

TL;DR

CoFi-Dec tackles LVLM hallucinations by integrating coarse-to-fine visual conditioning with a training-free decoding loop. It generates parallel textual hypotheses from original, coarse-grained, and fine-grained views, synthesizes corresponding images via a text-to-image model, and fuses three conditional token distributions at each step through a Wasserstein barycenter to enforce geometric consistency in the decoding trajectory. The approach yields substantial reductions in both object- and semantic-level hallucinations across six benchmarks and three open LVLMs, with strong performance on CHAIR, MME-Hallucination, and related tasks, while incurring a moderate inference-time overhead. This training-free, model-agnostic framework enhances grounding and faithfulness in multimodal generation, offering a practical path toward more reliable LVLM deployment in real-world settings.

Abstract

Large Vision-Language Models (LVLMs) have achieved impressive progress in multi-modal understanding and generation. However, they still tend to produce hallucinated content that is inconsistent with the visual input, which limits their reliability in real-world applications. We propose \textbf{CoFi-Dec}, a training-free decoding framework that mitigates hallucinations by integrating generative self-feedback with coarse-to-fine visual conditioning. Inspired by the human visual process from global scene perception to detailed inspection, CoFi-Dec first generates two intermediate textual responses conditioned on coarse- and fine-grained views of the original image. These responses are then transformed into synthetic images using a text-to-image model, forming multi-level visual hypotheses that enrich grounding cues. To unify the predictions from these multiple visual conditions, we introduce a Wasserstein-based fusion mechanism that aligns their predictive distributions into a geometrically consistent decoding trajectory. This principled fusion reconciles high-level semantic consistency with fine-grained visual grounding, leading to more robust and faithful outputs. Extensive experiments on six hallucination-focused benchmarks show that CoFi-Dec substantially reduces both entity-level and semantic-level hallucinations, outperforming existing decoding strategies. The framework is model-agnostic, requires no additional training, and can be seamlessly applied to a wide range of LVLMs. The implementation is available at https://github.com/AI-Researcher-Team/CoFi-Dec.

Figures (13)

• Figure 1: The illustration of the importance of combining coarse-grained and fine-grained visual information to avoid misleading global interpretations. By decomposing the scene, we uncover that the perceived red sunglasses are actually traffic lights, highlighting the need for multi-scale reasoning in visual understanding.
• Figure 1: Results on MMVP tong2024eyes. We apply our approach to LLaVA-1.5 liu2024improved and compare its performance against other hallucination mitigation methods.
• Figure 2: Overview of our proposed CoFi-Dec. Our method follows a coarse-to-fine generative feedback framework to enhance the fidelity of image-grounded text generation. By generating both coarse- and fine-grained textual descriptions from the original image and synthesizing corresponding images, we obtain multi-perspective visual feedback. These signals are then fused using a Wasserstein barycenter to produce a final response that is semantically rich, visually grounded, and resistant to hallucination.
• Figure 3: Case study on the LLaVA-Bench benchmark. Responses from standard decoding and our method (LLaVA-1.5) are shown with GPT-4V-assisted evaluations. Hallucinated and correct contents are highlighted in red and green, respectively.
• Figure A2: Effects of different generative models. We report the performance of different variants of our method, utilizing various stable diffusion models, on the LLaVA-1.5 backbone.

Theorems & Definitions (1)

• Remark 3.1