Published December 14, 2025
| Version v1
Preprint
Open
A Reasoning Critique of Diffusion Models
Authors/Creators
Description
We present a fundamental critique of diffusion models (particularly Diffusion Transformers)
as reasoning systems, grounded not in architectural specifics but in geometric necessity. Our
critique unfolds through three interconnected arguments:
First Critique (Ontological): We establish that diffusion models are Markovian denois
ing samplers, not sequential reasoning systems. This ontological classification has irreversible
consequences: memoryless sampling, stationary convergence to equilibrium πθ, and monotonic
information loss. Each denoising step erases history rather than accumulating state.
Second Critique (Dynamical): Building on the Markovian structure, we prove diffusion
training performs 1-shot sampling repeated n times, not chained causal modeling. Unlike
causal LMs where p(x1, . . . , xT |c) = Q t p(xt|x<t, c) chains dependent decisions, diffusion opti
mizes P t E[∥fθ(xt, c, t) − x0∥ 2 ]—n independent predictions. This is structural equivalence to
teacher forcing, inheriting its exposure bias and performance lower bound.
Third Critique (Topological): We prove that 1-shot repetition structure causes topolog
ical collapse in representation space. RNN-style chained reasoning builds cumulative structure
through ht+1 = ht +F, creating branching trajectories and unreachable holes (A = A∗ , Yonglin
Formula). But 1-shot repetition performs n independent predictions—no cumulative structure,
no branching. Result: manifolds collapse to equilibrium dead water (883M× flatter, 3.49× more
uniform, zero holes vs. Euler-Stack’s 400).
The Chain: Markovian sampler (1) → 1-shot repeated n times (2) → topological collapse
to dead water (3). Each critique builds on the previous, forming an unbreakable logical chain.
Verdict: Diffusion models are stationary samplers, not reasoning systems—appropriate
for generative tasks but fundamentally incapable of dynamic causal inference. This is not
circumvention; it is architectural regression.