The authors study exponential tilting: reweighting a base distribution so that outcomes with higher rewards become more probable.
Shifting denoiser inputs enables guided sampling from discrete diffusion models
McDonald and Bartlett derive exact identities for reusing a base denoiser to sample tilted discrete distributions, and modify training to cover the required inputs.
Big Tech
Curtis McDonald · Peter Bartlett
Simons Institute · University of California Berkeley · Google DeepMind
Research Digest··2 min read
Curtis McDonald and Peter Bartlett propose guiding a diffusion model by changing where its denoiser is evaluated, rather than adding a correction to its output.
§