Taming Object Hallucinations with Verified Atomic Confidence Estimation

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Liu, Jiarui, Xuan, Weihao, Jin, Zhijing, Diab, Mona
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866914153465118720
author Liu, Jiarui
Xuan, Weihao
Jin, Zhijing
Diab, Mona
author_facet Liu, Jiarui
Xuan, Weihao
Jin, Zhijing
Diab, Mona
contents Multimodal Large Language Models (MLLMs) often suffer from hallucinations, particularly errors in object existence, attributes, or relations, which undermine their reliability. We introduce TACO (Verified Atomic Confidence Estimation), a simple framework that mitigates hallucinations through self-verification and confidence calibration without relying on external vision experts. TACO decomposes responses into atomic queries, paraphrases them to reduce sensitivity to wording, and estimates confidence using self-consistency (black-box) or self-confidence (gray-box) aggregation, before refining answers with a language model. Experiments on five benchmarks (POPE, MME, HallusionBench, AMBER, and MM-Hal Bench) with two MLLMs (\texttt{LLaVA-1.5-7B} and \texttt{CogVLM2}) show that TACO consistently outperforms direct prompting and Visual Contrastive Decoding, reduces systematic biases, and improves confidence calibration, demonstrating its effectiveness in enhancing the faithfulness of MLLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2511_09228
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Taming Object Hallucinations with Verified Atomic Confidence Estimation
Liu, Jiarui
Xuan, Weihao
Jin, Zhijing
Diab, Mona
Computer Vision and Pattern Recognition
Computation and Language
Multimodal Large Language Models (MLLMs) often suffer from hallucinations, particularly errors in object existence, attributes, or relations, which undermine their reliability. We introduce TACO (Verified Atomic Confidence Estimation), a simple framework that mitigates hallucinations through self-verification and confidence calibration without relying on external vision experts. TACO decomposes responses into atomic queries, paraphrases them to reduce sensitivity to wording, and estimates confidence using self-consistency (black-box) or self-confidence (gray-box) aggregation, before refining answers with a language model. Experiments on five benchmarks (POPE, MME, HallusionBench, AMBER, and MM-Hal Bench) with two MLLMs (\texttt{LLaVA-1.5-7B} and \texttt{CogVLM2}) show that TACO consistently outperforms direct prompting and Visual Contrastive Decoding, reduces systematic biases, and improves confidence calibration, demonstrating its effectiveness in enhancing the faithfulness of MLLMs.
title Taming Object Hallucinations with Verified Atomic Confidence Estimation
topic Computer Vision and Pattern Recognition
Computation and Language
url https://arxiv.org/abs/2511.09228