Taming Object Hallucinations with Verified Atomic Confidence Estimation

TL;DR

TACO is a framework that reduces hallucinations in multimodal large language models by decomposing responses into atomic queries, estimating confidence through self-verification, and refining answers, thereby improving reliability and calibration.

Contribution

TACO introduces a novel self-verification and confidence calibration framework that enhances the faithfulness of MLLMs without external vision experts.

Findings

Outperforms direct prompting and Visual Contrastive Decoding

Reduces systematic biases in MLLMs

Improves confidence calibration across benchmarks

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 (\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.