Multi-Agent Intelligence for Multidisciplinary Decision-Making in Gastrointestinal Oncology

By: Rongzhao Zhang · Junqiao Wang · Shuyun Yang · Mouxiao Bian · Chao Ding · Yuwei Bai · Chihao Zhang · Yuguang Shen · Lei Wang · Lei Zheng · Qiujuan Yan · Yun Zhong · Meiling Liu · Jiwei Yu · Zheng Wang · Jie Xu · Meng Luo

10 December 2025 at 13:00

arXiv:2512.08674v1 Announce Type: new Abstract: Multimodal clinical reasoning in the field of gastrointestinal (GI) oncology necessitates the integrated interpretation of endoscopic imagery, radiological data, and biochemical markers. Despite the evident potential exhibited by Multimodal Large Language Models (MLLMs), they frequently encounter challenges such as context dilution and hallucination when confronted with intricate, heterogeneous medical histories. In order to address these limitations, a hierarchical Multi-Agent Framework is proposed, which emulates the collaborative workflow of a human Multidisciplinary Team (MDT). The system attained a composite expert evaluation score of 4.60/5.00, thereby demonstrating a substantial improvement over the monolithic baseline. It is noteworthy that the agent-based architecture yielded the most substantial enhancements in reasoning logic and medical accuracy. The findings indicate that mimetic, agent-based collaboration provides a scalable, interpretable, and clinically robust paradigm for automated decision support in oncology.

NeuroGenPoisoning: Neuron-Guided Attacks on Retrieval-Augmented Generation of LLM via Genetic Optimization of External Knowledge

cs.AI, q-bio.NC updates on arXiv.org

By: Hanyu Zhu · Lance Fiondella · Jiawei Yuan · Kai Zeng · Long Jiao

27 October 2025 at 12:00

arXiv:2510.21144v1 Announce Type: new Abstract: Retrieval-Augmented Generation (RAG) empowers Large Language Models (LLMs) to dynamically integrate external knowledge during inference, improving their factual accuracy and adaptability. However, adversaries can inject poisoned external knowledge to override the model's internal memory. While existing attacks iteratively manipulate retrieval content or prompt structure of RAG, they largely ignore the model's internal representation dynamics and neuron-level sensitivities. The underlying mechanism of RAG poisoning has not been fully studied and the effect of knowledge conflict with strong parametric knowledge in RAG is not considered. In this work, we propose NeuroGenPoisoning, a novel attack framework that generates adversarial external knowledge in RAG guided by LLM internal neuron attribution and genetic optimization. Our method first identifies a set of Poison-Responsive Neurons whose activation strongly correlates with contextual poisoning knowledge. We then employ a genetic algorithm to evolve adversarial passages that maximally activate these neurons. Crucially, our framework enables massive-scale generation of effective poisoned RAG knowledge by identifying and reusing promising but initially unsuccessful external knowledge variants via observed attribution signals. At the same time, Poison-Responsive Neurons guided poisoning can effectively resolves knowledge conflict. Experimental results across models and datasets demonstrate consistently achieving high Population Overwrite Success Rate (POSR) of over 90% while preserving fluency. Empirical evidence shows that our method effectively resolves knowledge conflict.

A Novel Urine DNA Predictor for Noninvasive Early Diagnosis and Monitoring Minimal Residual Disease of Upper Tract Urothelial Carcinoma

MRD

By: Wei Zuo · Xuanjun Guo · Jilong Zhang · Wei Yu · Yanrui Zhang · Huanqing Cheng · Qi Tang · Xuesong Li · Zhisong He · Liqun Zhou

23 October 2024 at 18:00

Cancer Med. 2024 Oct;13(20):e70346. doi: 10.1002/cam4.70346.

ABSTRACT

BACKGROUND: For early detection and postoperative monitoring of upper tract urothelial carcinoma (UTUC), the traditional detection method was limited to its invasiveness and insufficient sensitivity. We aim to use urine tumour DNA (utDNA) for detecting minimal residual disease (MRD), early diagnosis and perioperative monitoring in UTUC.

METHOD: We previously established a utDNA multidimensional bioinformatic valuation model, named utLIFE, using low-coverage whole-genome sequencing and targeted deep sequencing. This prospective cohort enrolled 93 patients diagnosed with UTUC without metastasis. We collected morning urine samples on the day of surgery and the discharge day after the operation for utLIFE testing. In addition, we also enrolled 80 healthy controls to further validate the specificity of the utLIFE model in the study.

RESULTS: The utLIFE of preoperative samples could discriminate UTUC with high specificity (96.25%, 77/80), and high sensitivity (96.77%, 90/93) regardless of stage and grade. The sensitivity of utLIFE was significantly higher than urine cytology (p < 0.001) and fluorescence in situ hybridisation (FISH) (p < 0.001) (N = 19), especially in early-stage and low-grade UTUC. Postoperative utLIFE scores were significantly decreased compared with those of preoperative samples (79 vs. 36, p < 0.001), indicating its association with tumour burden. For special pathology types, utLIFE performed less well in sensitivity and perioperative alteration.

CONCLUSION: In conclusion, we established a bioinformatic utDNA valuation model, utLIFE, which was validated to be a rapid and noninvasive approach with high sensitivity for early detection and MRD monitoring for UTUC.

PMID:39440792 | PMC:PMC11497171 | DOI:10.1002/cam4.70346

Reading view