HoliGround: Holistic Assessment for Grounded Chain-of-Thought

Paper

Abstract

Grounded Chain-of-Thought (GCoT) has emerged as a promising approach to promote the use of Multimodal Large Language Models (MLLMs) in critical applications by bringing interpretability and mitigating visual hallucinations. However, the faithfulness of these intermediate rationales is rarely assessed, as current evaluation protocols focus almost exclusively on final answer accuracy. This divergence can translate into real harm if users over-trust unfaithful rationales. Thus, we must establish evaluation protocols that move beyond simple accuracy to provide stronger empirical validation of a model's faithfulness, ensuring it is correct for the right reasons. To bridge this critical gap, we introduce HoliGround, a comprehensive benchmark designed to evaluate GCoT-capable models beyond their final predictions. We propose a novel, low-constraint metric that aggregates three fundamental dimensions: Logic Quality, which assesses the relevance and ordinal consistency of the reasoning sequence; Complexity Awareness, which weights evaluation based on step-level visual task difficulty; and Holistic Assessment, which integrates intermediate grounding fidelity with final answer correctness. Furthermore, we develop a deterministic, rule-based synthesis pipeline to construct HoliGround, offering a highly cost-effective and scalable method for generating GCoT annotations. By releasing both the benchmark and our synthesis framework, HoliGround establishes a new standard for rigorously validating the underlying reasoning of multimodal models.