Apple research shows confidence-based discrete diffusion samplers hit theoretical limits
Apple ML Research · rss · 2026-10-02
Apple ML Research's paper "Limits of Confidence in Diffusion" analyzes discrete diffusion models, including remasking and uniform-state samplers, which write multiple token positions per step drawn from per-position distributions.
The authors prove that a step matches the training distribution only when the written positions are conditionally independent given the already-fixed tokens, and that no product of per-position distributions can match a dependent group of tokens. Since pixels, phonemes, and words exhibit inherent dependencies, confidence-based sampling strategies face fundamental limitations in these domains.
More from Research
- Why AI can't solve the mystery of time: training presupposes the very clock it must explain — johnseach · 2026-10-03
- arXiv trends suggest AI uplift hits quantum physics within a year, all physics by late 2029 — cephaloform · 2026-10-03
- Berkeley's Humanoid Intelligence Center wins both tracks at IROS 2026 RoCo Challenge — berkeley_ai · 2026-10-03
- OpenStamp embeds watermarks into open-source LLM weights so users can't strip them — danish037 · 2026-10-03
- CruxBench lands at NeurIPS: frontier LLMs barely beat random at asking the right questions — mengyer · 2026-10-03
- Lampinen calls out neurosymbolic camp for walking back internal-symbolism claims — AndrewLampinen · 2026-10-03