Mitigating Batch Effects in Histopathology via Language-Mediated Robust Embedding Generation

A new framework called GLMP aims to reduce batch effects in computational pathology by converting histology image patches into text descriptions before generating numerical embeddings, according to research submitted in 2026 [1][2]. Pathology foundation models, or PFMs, have shown strong potential in clinical and scientific tasks, but their performance is frequently undermined by batch effects — non-biological variations introduced by different tissue source institutions, known as TSIs [1][2]. These artifacts distort learned feature representations and weaken a model’s ability to generalize across hospitals and laboratories. Conventional countermeasures such as stain normalization have provided limited success against these high-dimensional, complex distortions [1][2]. The proposed model, called the General-purpose LLM-Mediated Pathology model (GLMP), takes a different route. It first uses a general-purpose multimodal large language model, or MLLM, to produce a textual description of the histological features visible in an image patch [1][2]. That intermediate text is then encoded into a numerical representation by a text encoder. The authors state that GLMP is the first pathology model to use text descriptions of histological features as an intermediate representation for generating numerical embeddings from histology images [1][2]. By prioritizing biologically meaningful signals over TSI-specific artifacts, the framework is designed to improve cross-institutional generalization [1][2]. Large language models, the broader class of technology underpinning the text-based step in GLMP, are machine learning models with many parameters trained on vast amounts of text through self-supervised learning [8]. The GLMP paper highlights what it calls the untapped potential of broad-domain, non-specialized MLLMs in computational pathology [1][2]. The research was posted on arXiv, the open-access repository that hosts preprints in fields such as computer science, statistics, and electrical engineering [4]. arXiv has expanded its offerings through arXivLabs, a framework that lets collaborators build and share new features directly on the platform [1]. One such collaboration is with Hugging Face, which since October 2021 has enabled the community to create and share over 12,000 open-source machine learning demos through Hugging Face Spaces [3]. These demos, built with tools such as Gradio and Streamlit, allow users to interact with models in a browser without writing code [3][4]. Authors can link a Space to their arXiv paper by including the paper’s identifier in the Space’s README file or by associating the Space with a model hosted on the Hugging Face Hub [5]. The integration places a Demos tab on the paper’s abstract page, making it possible for readers to test models immediately [3][4][5].