Talk: Eliciting Any-Order Inference from Any-Order Models

NandoDF · x · 2026-09-01

Links to a recording of the talk 'From Interface to Inference: Eliciting Any-Order Inference from Any-Order Models'. It discusses non-causal discrete reasoning tasks like code generation, where programmers move between structure and details (any-order inference). It highlights core limitations of masked diffusion models: the interface doesn't guarantee inference capability, and fixed-canvas token-level models suffer from 'positional uncertainty'.

Original post →

More from Research

Research channel →