Time-Anchored Diffusion Language Models: Latent-Space Caching for Fast Generation¶

L1 mechanism diagnostic. Stale-anchor correction, without loading DiffusionGemma-26B.

Paper information¶

Field	Value
Paper	arXiv 2609.37924
Affiliations	The University of Texas at Austin, first author's affiliation
First public release	2026-09-29 (arXiv v1)
Authors' code	No public author repository found as of 2026-10-01.
Adapter / method	`tadm`
Local code	`src/auto_research/foundation_latest_20261001.py`

Original paper¶

Background and contribution¶

Cache an expensive anchor network's latent representation across reverse-diffusion steps and refresh it periodically. A small fusion module uses the current state to gate a correction before the lightweight denoiser.

flowchart LR
  A[Cached anchor] --> F[Gated correction] --> D[Lightweight denoiser]
  C[Current state] --> F

Key figure from the paper¶

Original Figure 1: anchor caching and correction. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\(\tilde h_{t|t'}=h_{t'}+g_\phi(c_t,h_{t'})\odot\Delta_\phi(c_t,h_{t'})\). The local component implements the gated correction.

Paper results¶

The post-trained DiffusionGemma-26B variant reports roughly 49%–79% more throughput. The pretrained version reduces layer computation by up to 38% and reports up to 73% more throughput than ADLM.

Local reproduction¶

Compare the disabled/default mechanism with the enabled one. Three-seed metrics are diagnostic_only=true, checking transitions, gradients, or scheduling, not capability or cross-model improvement.

Reproduction limits¶

No DiffusionGemma-26B checkpoint run; these results do not reproduce paper throughput.