Math 2.0

Resources / preprint

Diffusion Transformers are Provably Optimal In-context Generators N37WKFTR7Q

URIm2://res/N37WKFTR7Q
Kindpreprint
CreatorsGuoji Fu, Tomoya Wakayama, Ryotaro Kawata, Atsushi Nitanda, Wee Sun Lee, Taiji Suzuki
Date2026-10-04
Subjectsarxiv cs.LG, arxiv math.ST, arxiv stat.ML, arxiv stat.TH
Scopeborder

Expressions

e1: version, preprint

Date2026-10-04
IdsarXiv 2610.05333

Items in this resource

The mentions of the inventory that name this resource. Each name links to the card of its inventory row.

ItemKindWhereSource
Token bounds and parameter sensitivity of the four-attention-block networkstatementlemma E.8arxiv
Episode-level KL control of integrated score errorstatementproposition D.4arxiv
Clipping error for the scaled scorestatementlemma E.5arxiv
Finite-episode score learning with variance-balanced DSMstatementtheorem F.2arxiv
Uniform upper and matching lower bounds for finite-prompt task uncertaintystatementtheorem H.1arxiv
Randomized time-grid bound for at most $M$ updatesstatementcorollary G.6arxiv
Transformer size and covering numbersstatementproposition E.7arxiv
Sensitivity of scaled dot-product attentionstatementlemma E.2arxiv
8PHPBXT6KCstatementcorollary 3.3arxiv
Exact reverse-score dynamics on a positive-time intervalstatementlemma G.4arxiv
Conditional early-stopping error boundstatementcorollary G.2arxiv
Scaled-score bound and truncation errorstatementlemma D.1arxiv
Likelihood and conditional-density regularitystatementlemma D.2arxiv
Normalized exponential familiesexampleexample I.5arxiv
H\"older early-stopping error bound on the cubestatementlemma G.1arxiv
Normalized OU-basis regularitystatementlemma D.3arxiv
Fixed-dimensional H\"older approximation by ReLU networksstatementlemma E.3arxiv
Prompt-generating processdefinitiondefinition 2.1arxiv
Prompt-conditioned score Transformerdefinitiondefinition E.1arxiv
Smoothly varying multimodal mixturesexampleexample I.4arxiv
a latent KL modulus to one-step predictionstatementlemma H.3arxiv
Posterior averaging boundstatementlemma H.2arxiv
Gaussian-tail bounds for the selected radiistatementlemma G.7arxiv
Four-block Transformer constructionstatementlemma E.1arxiv
N2ES9GHS6Mstatementtheorem G.8arxiv
Time-local OU-basis approximationstatementlemma E.4arxiv
Q9VNP43EB1statementtheorem 3.2arxiv
Smooth perturbations of the uniform densityexampleexample I.3arxiv
Finite-energy KL bound for diffusions with matched noisestatementlemma G.3arxiv
Randomized time-grid discretization boundstatementproposition G.5arxiv
SHG3KRBZ8Mstatementtheorem 3.1arxiv
Integrated score approximation by the Transformer classstatementproposition E.6arxiv
Smoothly varying conditional output densitiesexampleexample I.2arxiv
Smoothly varying sample densitiesexampleexample I.1arxiv
Exact variance-balanced DSM excess-risk identitystatementproposition C.1arxiv
Bias from clipping Gaussian DSM labelsstatementproposition F.1arxiv
Task-conditional generation at the sharp prompt ratestatementcorollary H.4arxiv

Catalogue records