生成但驗證：透過回顧性重採樣降低視覺語言模型中的幻覺現象

Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling

April 17, 2025

作者: Tsung-Han Wu, Heekyung Lee, Jiaxin Ge, Joseph E. Gonzalez, Trevor Darrell, David M. Chan

cs.AI

摘要

視覺語言模型（VLMs）在視覺理解方面表現卓越，但常常面臨視覺幻覺問題，即生成對不存在物體、動作或概念的描述，這在安全關鍵應用中構成重大風險。現有的幻覺緩解方法通常遵循兩種範式之一：生成調整，即修改解碼行為以使文本與視覺輸入對齊；以及事後驗證，即使用外部模型評估並修正輸出。雖然有效，但生成調整方法通常依賴啟發式方法且缺乏修正機制，而事後驗證則較為複雜，通常需要多個模型且傾向於拒絕輸出而非精煉它們。在本研究中，我們引入了REVERSE，一個將幻覺感知訓練與即時自我驗證相結合的統一框架。通過利用包含超過130萬半合成樣本的新幻覺驗證數據集，以及一種新穎的推理時回顧性重採樣技術，我們的方法使VLMs能夠在生成過程中檢測幻覺並動態修正這些幻覺。我們的評估顯示，REVERSE在幻覺減少方面達到了最先進的水平，在CHAIR-MSCOCO和HaloQuest上分別比現有最佳方法高出12%和28%。我們的數據集、模型和代碼可在以下網址獲取：https://reverse-vlm.github.io。

English

Vision-Language Models (VLMs) excel at visual understanding but often suffer from visual hallucinations, where they generate descriptions of nonexistent objects, actions, or concepts, posing significant risks in safety-critical applications. Existing hallucination mitigation methods typically follow one of two paradigms: generation adjustment, which modifies decoding behavior to align text with visual inputs, and post-hoc verification, where external models assess and correct outputs. While effective, generation adjustment methods often rely on heuristics and lack correction mechanisms, while post-hoc verification is complicated, typically requiring multiple models and tending to reject outputs rather than refine them. In this work, we introduce REVERSE, a unified framework that integrates hallucination-aware training with on-the-fly self-verification. By leveraging a new hallucination-verification dataset containing over 1.3M semi-synthetic samples, along with a novel inference-time retrospective resampling technique, our approach enables VLMs to both detect hallucinations during generation and dynamically revise those hallucinations. Our evaluations show that REVERSE achieves state-of-the-art hallucination reduction, outperforming the best existing methods by up to 12% on CHAIR-MSCOCO and 28% on HaloQuest. Our dataset, model, and code are available at: https://reverse-vlm.github.io.