OncoTriad-QA: A New Benchmark for Pan-Cancer Multimodal Reasoning
OncoTriad-QA is a major leap forward in how we evaluate medical large language models (LLMs) and vision-language models (VLMs) for oncology at the patient level. It stands out by merging data from radiology, pathology, genomics, and clinical metadata, addressing a gap that exists in current benchmarks that usually focus on individual areas. This benchmark includes 86.1k semantic questions from 9,281 TCGA patient cases across 32 cancer types. It pulls together information from CT/MRI scans, whole-slide histopathology, genetic mutations, copy-number changes, DNA methylation, bulk RNA sequencing, and clinical data. Annotations are created through a source-grounded pipeline that leverages curated labels, diagnostic reports, and molecular information, aiming to assess pan-cancer reasoning for medical AI applications.
Key facts
- OncoTriad-QA is a patient-level radiology-pathology-genomics benchmark for pan-cancer question answering.
- It contains 86.1k semantic questions across 9,281 TCGA patient cases from 32 cancer cohorts.
- The benchmark aligns CT/MRI radiology, whole-slide histopathology, somatic mutations, copy-number alterations, DNA methylation, bulk RNA-seq, and clinical metadata.
- Annotations are constructed through a source-grounded LLM-assisted pipeline using curated labels, diagnostic reports, molecular profiles, and modality-derived evidence.
- Most medical LLM and VLM benchmarks focus on isolated modalities or narrow image-text tasks, leaving patient-level oncology assessment untested.
- The benchmark is introduced to address the gap in evaluating patient-level oncology assessment across multiple evidence streams.
- The source is an arXiv paper with ID 2608.02615.
- The paper is announced as a cross-type submission.
Entities
Institutions
- arXiv
- TCGA