README.md

Learning Unmasking Policies Reproduction

CPU-only evidence package for Learning Unmasking Policies for Diffusion Language Models.

The package audits the public Apple repository and performs deterministic local checks of masked-state transitions, confidence-based token selection, block schedules, and expert left-to-right unmasking. Benchmark claims that require trained checkpoints or raw evaluation outputs are marked inconclusive unless those artifacts are present.

Run

uv run --project submissions/learning-unmasking-policies-for-diffusion-language-models python submissions/learning-unmasking-policies-for-diffusion-language-models/generate_evidence.py
uv run --project submissions/learning-unmasking-policies-for-diffusion-language-models python -m pytest submissions/learning-unmasking-policies-for-diffusion-language-models/tests -q