# Preserving Mathematical Reasoning in Compressed Diffusion Language Models via Trajectory-Aware Low-Rank Approximation

Source: https://arxiv.org/abs/2610.03326
Published: 2026-10-02
Status: unreviewed source report

## Source excerpt (up to 25 words)

> Diffusion language model (dLLM) compression faces a known challenge because calibration is typically performed on clean, fully visible activations, whereas inference traverses partially masked intermediate

## Review question

Inspect the formal statement, imported axioms, admitted lemmas, checker version and correspondence with the informal theorem. A successful compiler exit alone does not prove that the intended question was formalized.

No independent proof check or claim of peer review.