Measure every stage
Evaluate original inputs, codec reconstructions, latent generations, and decoded outputs under comparable protocols.
Discrete latent generation
How can a generative model compress its state without hiding where quality is lost—or what should remain fixed?
Research question
This direction studies discrete intermediate representations as both a computational bottleneck and a control surface. One line of work diagnoses whether quality loss originates in the codec or the generator. Another asks whether coarse latent positions can expose inspectable choices about what a software generator preserves and what it may change.
Working principles
Separate proxy from outcome
Codebook usage or latent geometry is diagnostic evidence, not proof that decoded text or code improved.
Expose the control surface
Representations should make preservation and change scope inspectable rather than leaving them implicit in a prompt.