PathoSage: Building Trustworthy Patch-Level AI in Pathology

Thesis

Current multimodal pathology models often generate convincing but inaccurate descriptions of tissue morphology, making them risky for clinical use. PathoSage proposes a workflow that treats each piece of evidence as a separate entity, then adjudicates between them with an experience‑aware agent, aiming to reduce hallucinations and improve patch‑level reliability.

Evidence

The arXiv preprint released on June 9, 2026 notes that end‑to‑end pathology large language models (LLMs) frequently hallucinate morphological features. It also points out that existing agentic systems blend tool outputs and retrieved knowledge into a single context, leaving decisions exposed to contradictory evidence and context contamination. PathoSage’s core contribution is an architecture that keeps sources distinct and lets an experience‑aware agent weigh them before reaching a conclusion.

Context

Multimodal LLMs have shown promise in interpreting whole‑slide images, yet their black‑box nature makes it hard to trust individual patch predictions. In parallel, OpenAI’s recent Codex rollout (June 2, 2026) highlights a trend toward role‑specific plugins that extend AI capabilities without deep model changes. NVIDIA’s June 3, 2026 announcements about physical AI agents for robotics and autonomous driving illustrate a broader industry push to embed AI agents in concrete workflows, where safety and repeatability are non‑negotiable. PathoSage sits at the intersection of these trends: it adopts an agentic mindset while focusing on the evidential rigor needed for pathology.

Counter‑Arguments

Critics may argue that separating evidence streams adds latency and computational overhead, especially when processing high‑resolution whole‑slide images. The same paper that introduces PathoSage admits that merging tool outputs is currently the default because it simplifies pipeline design. Additionally, the experience‑aware component relies on prior interactions, which could bias the system toward familiar patterns and overlook novel pathology presentations.

Prediction

If PathoSage’s adjudication mechanism proves scalable, it could become a template for other high‑stakes domains—radiology, genomics, even autonomous vehicle perception—where conflicting sensor data must be reconciled. Builders may adopt a modular evidence‑first pattern: retrieve, isolate, and then let an agent decide, rather than feeding everything into a monolithic model. Success will likely hinge on open tools that let developers plug in custom evidence sources without re‑training the core model, echoing the flexibility championed by OpenAI’s Codex plugins and NVIDIA’s agent‑skill framework.