Trust but Verify: Designing Region-Inhomogeneous Forward Dynamics for Discrete Diffusion

Trust but Verify reopens the design space for discrete diffusion by combining absorbing and uniform-mixing kernels with region-dependent rates, rather than using binary trust through absorbing corruption. Missing positions follow absorbing dynamics while context undergoes low-rate random substitution, exposing the model to observed but uncertain tokens during training. At inference, a heuristic sampler re-injects noise into suspicious context under an annealed intensity and freezes context updates late in the trajectory.