level: research
deterministic few-step generation works well on continuous image latents but produces incoherent text on continuous text latents. the problem is not a training or scaling issue. it is geometric. a smooth, regularity-limited deterministic map cannot resolve a discrete branch choice before a sharp categorical readout. so few-step failure is governed by decoder sharpness, not transport accuracy.
in the overlapping regime of real text autoencoders, a theorem proves that the posterior-mean terminal step flips tokens at a rate equal to the latent mass in a tube around decision boundaries. the tube width scales with the decoder's sharpness. two diagnostics measure this: dabi for readout sharpness and cci for categorical commitment. measurements on published checkpoints show that four independent continuous-text decoders amplify a boundary-aligned perturbation far beyond a norm-matched isotropic one. dabi values range from 500 to over 100,000. image decoders have dabi around 1.
the findings explain why text latents need many steps or stochastic sampling. the sharp decision boundaries in text decoders make deterministic few-step mapping unreliable. this geometric insight can guide the design of better text autoencoders. reducing decoder sharpness or adding stochasticity could enable faster text generation. the diagnostics dabi and cci provide practical tools to evaluate and compare decoder behavior.
why it matters: understanding this geometric limitation helps improve fast text generation models by focusing on decoder sharpness instead of just scaling or training.