PathGLS: Evaluating Pathology Vision-Language Models without Ground Truth through Multi-Dimensional Consistency

Dimensions overall score 9.0

• PathGLS reveals a 40.2% sensitivity drop for hallucinated reports on the Quilt-1M dataset compared to only 2.1% for BERTScore.result 95%
• PathGLS achieves a Spearman's rank correlation of ρ=0.71 (p < 0.0001) when validated against expert-defined clinical error hierarchies.result 95%
• PathGLS significantly outperforms LLM-based approaches like Gemini 3.0 Pro (ρ=0.39, p < 0.0001) in correlation with clinical error hierarchies.result 90%
• PathGLS is a reference-free evaluation framework that assesses pathology VLMs across three dimensions: Grounding, Logic, and Stability.method 98%
• PathGLS supports both patch-level and whole-slide image (WSI)-level analysis, yielding a comprehensive trust score.technical 95%
• Experiments demonstrate the superiority of PathGLS on Quilt-1M, TCGA, REG2025, PathMMU and TCGA-Sarcoma datasets.result 90%
• PathGLS serves as a reliable criterion for benchmarking VLMs on private clinical datasets and informing safe deployment.market 85%
• PathGLS directly quantifies hallucination rates and domain shift robustness in pathology VLMs.technical 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card