Latent Research Landscapes

I embedded the SIGGRAPH Asia archive with OpenAI's ada embeddings, one vector for each paper abstract. The landscape below is a 2D projection of that space, with kernel density estimation raising peaks where the abstracts cluster.

Four density landscapes, each with vertical pillars and a generated research abstract and title — Each blob is a region of the SIGGRAPH Asia embedding space. The pillars are the real papers nearest a click. The abstract and title are generated from their blend, not real.

Clicking a point on the map selects the abstract embeddings nearest to it, the pillars in the image. Blending those vectors together gives a new point that sits between real papers but belongs to none of them.

I run that blended vector back through vec2text, an embedding reversal model that decodes a vector into text. So the loop is: embed real abstracts, blend them, and generate a new abstract and title for a paper that does not exist. Research read out of the latent space between the papers that do.