Every mainstream generative tool — image, video, text — has converged on the same interface: type a sentence, get an output, type another sentence to adjust it. A preprint posted to arXiv on October 27 argues that convergence happened by default, not because language is actually a good way to do fine-grained creative work, and proposes an alternative built around spatial navigation instead of prompting.
The tool, called PolyGandr, comes from a team that includes Ryan P. Adams (Princeton) and Remco Chang (Tufts, where he directs work on visual analytics and data visualization). Their argument: language is genuinely good at specifying high-level intent — “a cat in the style of Van Gogh” — but it’s inefficient and imprecise for the kind of iterative refinement and open-ended exploration an artist does once they’re already close to what they want. You can’t easily say “a little more like that one, but warmer” in a prompt box without starting over.
PolyGandr instead treats a generative model’s latent space — the high-dimensional internal representation the model samples from to produce outputs — as a literal space to move through. Rather than typing a new description each time, a user navigates relationships between points in that space visually, letting proximity and direction do the work that adjective-stacking does in a text prompt.
The name and framing follow an earlier tool from an overlapping part of the same research area, LatentGandr (posted to arXiv in April 2026), which built interactive visualizations of local neighborhoods inside a latent space by computing local principal components — a way of showing a user what directions of change are available from where they currently are, rather than making them guess. PolyGandr reads as an extension of that same underlying idea into a more general steering interface rather than a one-off visualization technique, which is worth knowing if you want to trace where this line of work is headed next.
The paper carries no accepted-venue listing yet, so this is an unreviewed preprint, not a peer-reviewed result — worth tracking as a direction rather than citing as settled. The practical test, as with every “better than prompting” interface claim, will be whether anyone outside the group that built it can pick it up and actually get more precise control out of it than a well-written prompt already gives them.