I wanted to share a simple RL environment I built that gave frontier LLMs a set of tools that lets it zoom and pan across a digitized pathology slide to find the relevant regions to make a diagnosis. Here are some videos of the LLM performing diagnosis on a few slides:
(https://www.youtube.com/watch?v=k7ixTWswT5c): traces of an LLM choosing different regions to view before making a diagnosis on a case of small-cell carcinoma of the lung
(https://youtube.com/watch?v=0cMbqLnKkGU): traces of an LLM choosing different regions to view before making a diagnosis on a case of benign fibroadenoma of the breast
Why I built this:
Pathology slides are the backbone of modern cancer diagnosis. Tissue from a biopsy is sliced, stained, and mounted on glass for a pathologist to examine abnormalities.
Today, many of these slides are digitized into whole-slide images (WSIs)in TIF or SVS format and are several gigabytes in size.
While there exists several pathology-focused AI models, I was curious to test whether frontier LLMs can perform well on pathology-based tasks. The main challenge is that WSIs are too large to fit into an LLM’s context window. The standard workaround, splitting them into thousands of smaller tiles, is inefficient for large frontier LLMs.
Inspired by how pathologists zoom and pan under a microscope, I built a set of tools that let LLMs control magnification and coordinates, viewing small regions at a time and deciding where to look next.
This ended up resulting in some interesting behaviors, and actually seemed to yield pretty good results with prompt engineering:
- GPT 5: explored up to ~30 regions before deciding (concurred with an expert pathologist on 4 out of 6 cancer subtyping tasks and 3 out of 5 IHC scoring tasks)
- Claude 4.5: Typically used 10–15 views but similar accuracy as GPT-5 (concurred with the pathologist on 3 out of 6 cancer subtyping tasks and 4 out of 5 IHC scoring tasks)
- Smaller models (GPT 4o, Claude 3.5 Haiku): examined ~8 frames and were less accurate overall (1 out of 6 cancer subtytping tasks and 1 out of 5 IHC scoring tasks)
Obviously, this was a small sample set, so we are working on creating a larger benchmark suite with more cases and types of tasks, but I thought this was cool that it even worked so I wanted to share with HN!