Simple Token-Level Confidence Improves Caption Correctness

TL;DR

This paper introduces Token-Level Confidence (TLC), a simple method to evaluate caption correctness in vision-language models, significantly improving accuracy in understanding and reducing hallucinations.

Contribution

The paper proposes TLC, a token-level confidence measure, that enhances caption correctness assessment and outperforms previous methods in various vision-language tasks.

Findings

TLC improves verb understanding accuracy by 10%.

TLC reduces object hallucination rates by 30%.

Learned confidence estimators further enhance performance.

Abstract

The ability to judge whether a caption correctly describes an image is a critical part of vision-language understanding. However, state-of-the-art models often misinterpret the correctness of fine-grained details, leading to errors in outputs such as hallucinating objects in generated captions or poor compositional reasoning. In this work, we explore Token-Level Confidence, or TLC, as a simple yet surprisingly effective method to assess caption correctness. Specifically, we fine-tune a vision-language model on image captioning, input an image and proposed caption to the model, and aggregate either algebraic or learned token confidences over words or sequences to estimate image-caption consistency. Compared to sequence-level scores from pretrained models, TLC with algebraic confidence measures achieves a relative improvement in accuracy by 10% on verb understanding in SVO-Probes and outperforms prior state-of-the-art in image and group scores for compositional reasoning in Winoground by a relative 37% and 9%, respectively. When training data are available, a learned confidence estimator provides further improved performance, reducing object hallucination rates in MS COCO Captions by a relative 30% over the original model and setting a new state-of-the-art.