Taming Object Hallucinations with Verified Atomic Confidence Estimation

2025-11-12

AI Summary

Multimodal large language models (MLLMs) frequently exhibit hallucinations in object existence, attribute, and relational reasoning, undermining output reliability. To address this, we propose TACO—a framework that decomposes model responses into verifiable atomic queries via **atomic query decomposition**, estimates black-box confidence through **sentence rewriting and self-consistency aggregation**, and enhances calibration via **gray-box language model fine-tuning**, all without external visual experts. TACO effectively mitigates systematic biases and consistently outperforms direct prompting and visual contrastive decoding across five benchmarks. It improves both faithfulness—measured by alignment between generated outputs and visual evidence—and confidence calibration accuracy. By enabling lightweight, general-purpose, and interpretable hallucination mitigation, TACO establishes a novel paradigm for enhancing MLLM reliability.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

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 ( exttt{LLaVA-1.5-7B} and exttt{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.

Problem

• Mitigating object hallucinations in Multimodal Large Language Models
• Reducing errors in object existence, attributes, and relations
• Improving confidence calibration without external vision experts

Innovation

• Decomposes responses into atomic queries for verification
• Paraphrases queries to reduce wording sensitivity
• Estimates confidence via self-consistency or self-confidence aggregation