Temporal Difference Learning for Diffusion Models (ICML 2026) arxiv.org/abs/2606.15048 By Yangchen Pan (my former PhD student) and co-authors. It reformulates diffusion training as a Markov reward process and introduces a TD objective to encourage temporal consistency across d…
AI Weekly's analysis
→
- The paper introduces a temporal difference objective that penalises inconsistency across the full denoising trajectory rather than only at adjacent time steps.
- It reframes diffusion as a Markov reward process and denoising as a policy evaluation problem, unifying discrete-time and continuous-time formulations.
- Reported FID gains are strongest when the number of sampling steps is small, the regime where few-step samplers and low-compute serving live.
Read full analysis →