level: technical
multimodal large language models often make mistakes when analyzing small regions of pathology images, sometimes inventing features that are not there. recent agent-based systems try to fix this by combining tool outputs and retrieved knowledge, but mixing all evidence into one context can lead to confusion and bias. pathosage tackles this by splitting the process into three clear stages: finding relevant knowledge, gathering evidence from different tools, and then judging that evidence separately.
the key part is structured evidence deliberation, which looks at each piece of evidence on its own, spots conflicts, and only then forms a final answer in a clean context. this avoids the problem of early information unfairly shaping the result. the system also includes a training-free experience module that tracks how reliable each tool has been over time, using a simple statistical model to weigh their contributions without needing extra training data.
by keeping evidence sources apart until the final decision, pathosage reduces hallucinations and makes patch-level reasoning more trustworthy. the approach does not require retraining the underlying models, making it easier to adopt in existing workflows. tests show it can handle conflicting information better than methods that dump everything into one prompt, pointing toward more reliable ai assistants for pathologists.
why it matters: this method can make ai pathology tools more reliable by reducing false findings, which is critical for real clinical use where mistakes can affect patient care.