1 Photos and videos
3 Oct 2025
🎉 Accepted to NeurIPS 2025!
A-CFG: Adaptive Classifier-Free Guidance
TL;DR: make the “unconditional” in CFG dynamic w.r.t. uncertainty.
Paper: arxiv.org/abs/2505.20199
#NeurIPS2025
2
8
3 Oct 2025
[2/n] What we do?
At each step:
- estimate per-token softmax confidence;
- re-mask low-confidence tokens;
build a localized unconditional, so CFG acts only where unsure.
6
3 Oct 2025
[1/n] Why?
Standard CFG uses a static null prompt / full-mask unconditional. In MDM/DLM, per-token confidence evolves over steps.
Static baselines can’t target truly uncertain positions → diluted guidance.
3