Abstract 6722: CUPAI-2: A collaborative multi-agent large language model framework for diagnosis of cancer of unknown primary.

Abstract Background: Cancers of unknown primary (CUP) accounts for 2-5% of cancers but remains a leading cause of death, as failure to identify the primary site blocks targeted therapy and trial access. We developed CUPAI-2, a multi-agent LLM framework with a website that integrates genomic data to predict CUP primary site. Methods: CUPAI-2 implements a Molecular Tumor Board (MTB)-inspired architecture comprising (i) domain specialist models and (ii) a meta-decision layer for consensus integration. The specialist tier comprises three LLMs (Qwen-2.5-0.5B, Qwen-3-0.6B, Llama-3.2-1B), each fine-tuned on modality specific corpora including single nucleotide variants, copy number alterations , and chromosomal translocations to produce independent primary site likelihoods and rationales. Their outputs (class probabilities, uncertainty scores, and rationale embeddings) feed a statistical coordinator that performs stacked generalization, employing an XGBoost model as a non-linear meta-learner to arbitrate and integrate predictions. A final oversight agent (DeepSeek) functions as the MTB manager, auditing discordant cases against structured genomic contexts and literature-derived priors, and issuing an adjudicated prediction. We analyzed 20,483 next generation sequencing (NGS) profiled tumors spanning 19 cancer types, and benchmarked CUPAI-2 against experienced oncologists on test sets. Results: CUPAI-2 achieved robust diagnostic performance with Top1 accuracy of 76.93% and Top3 of 92.82%. Performance exceeded that of oncologists on 190 uniformly distributed cancer cases (Top1: 71.05% vs. 68.42%; Top3: 83.15% vs. 80%), and inter-model reliability within the CUPAI-2 was high as reflected by Jaccard similarity coefficients. Compared with traditional machine learning (Top1/Top3: 60.38%/79.77%), CUPAI-2 achieved substantial improvements of 17.83 and 12.33 percentage points (pp) , respectively. It also delivered additional gains of ∼3-5 pp compared with non-collaborative, individually fine-tuned LLMs (75.27%/87.55%). CUPAI-2 further surpassed the current best published voting-based model, establishing a new state-of-the-art for CUP origin prediction while maintaining clinically acceptable computational efficiency. A prototype (https://cupai.origimed.com) enables uploading of NGS reports for CUP prediction and is not intended for clinical decision making. Conclusions: CUPAI-2, a coordinated MTB-inspired multi-agent LLM architecture, synthesizes heterogeneous genomic features to predict CUP primary sites with greater accuracy, calibration, and interpretability than individual LLMs, conventional models, or human experts. These findings highlight the feasibility and value of multi-agent LLM systems for high dimensional genomic reasoning and suggest a promising clinical decision making direction for improving cancer primary site inference. Citation Format: Junhan Zhao, Juejie Zhang, Lingjie Fan, Anzhi Chen, Zheyi Ji, Zanmei Xu, Yannan Zhu, Kai Wang. CUPAI-2: A collaborative multi-agent large language model framework for diagnosis of cancer of unknown primary abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6722.