This project is funded by the Alfred P. Sloan Foundation
in partnership with the United Kingdom Metascience Unit and Canada’s Social Sciences and Humanities Research Council
Florilegium of Latent Operations and Representations (FLOR)
Computational Poetics x Interpretability
Large language models belong to a long lineage of writing technologies, from the clay tokens of ancient Mesopotamia to the algorithmic tokens of today. Yet, compared to previous writing technologies, the materiality of LLMs is hard to engage. While it is possible to see the intermediate representations and activations of these systems, very little of this is legible to humans.
Interpretability research builds tools and techniques to read and make sense of the numbers inside LLMs. Interpretability researchers produce rich textual and graphical representations to understand how these models work, such as attention maps, the logit lens, and feature dashboards. For now, however, these tools are used and read largely by the computer science community.
Florilegium of Latent Operations and Representations (FLOR) brings interpretability research into dialogue with concrete and computational poetry. Its name comes from the medieval florilegium, a “gathering of flowers,” which referred to the practice of selecting excerpts and passages from authoritative texts and arranging them like a bouquet around a given topic. This practice enabled readers to reconfigure, question, and in some cases contest the texts they engaged with through the materials of manuscripts. Similarly, FLOR develops tools to collect and recompose the internal activations of LLMs to generate poetry.
Working with small open models, the project treats a model’s internals as poetic material. The aim is to open the tools of interpretability to textual scholars, poets, designers, and STS scholars, and to design margins for language models: surfaces that readers beyond the lab can write on.
Interpretability research builds tools and techniques to read and make sense of the numbers inside LLMs. Interpretability researchers produce rich textual and graphical representations to understand how these models work, such as attention maps, the logit lens, and feature dashboards. For now, however, these tools are used and read largely by the computer science community.
Florilegium of Latent Operations and Representations (FLOR) brings interpretability research into dialogue with concrete and computational poetry. Its name comes from the medieval florilegium, a “gathering of flowers,” which referred to the practice of selecting excerpts and passages from authoritative texts and arranging them like a bouquet around a given topic. This practice enabled readers to reconfigure, question, and in some cases contest the texts they engaged with through the materials of manuscripts. Similarly, FLOR develops tools to collect and recompose the internal activations of LLMs to generate poetry.
Working with small open models, the project treats a model’s internals as poetic material. The aim is to open the tools of interpretability to textual scholars, poets, designers, and STS scholars, and to design margins for language models: surfaces that readers beyond the lab can write on.
The project is currently being developed at LIRIS (CNRS) in Lyon and will continue at MIT’s Trope Tank starting in February 2027.
Drift, 2026. Ink on paper, pen plotter.
Six lines of T. S. Eliot’s “The Dry Salvages” (V), passed through GPT-2. At each word, the model assigns a probability to each possible next word. The larger line shows one word sampled from those probabilities at each position, creating a poem out of the model's distribution. The smaller line is Eliot’s original.