Yandex team proposes representation-space MMD as an evaluation metric for diffusion LMs
LucaAmb · x · 2026-10-27
A paper session on Nov 2 presents "Representation-Space MMD for Diffusion Language Models," presented by Dmitry Baranchuk (Yandex Research) and co-led by Ilya Drobyshevskiy and Ilia Sudakov, with authors from Yandex Research, HSE University, Applied AI Institute, and others.
The method maps reference and generated samples into representation features via a pretrained diffusion LM φ, then computes token-level RBF-kernel MMD (maximum mean discrepancy) between the two feature sets, providing a principled evaluation tool for diffusion language models.
More from Research
- Open model from Johns Hopkins and Stanford spots adrenal tumor on CT in one click — MaziyarPanahi · 2026-10-27
- Philosopher Seth Lazar launches OSPHAIC, an Oxford-era platform for philosophy of AI and computing — sethlazar · 2026-10-27
- Full receipts published for open re-check of Goodfire SAE steering claim — kevrussell · 2026-10-27
- Blind human check: rater picks steered image in 100 of 100 pairs in pre-registered SAE re-check — kevrussell · 2026-10-27
- Pre-registered re-run confirms Goodfire's SAE steering claim on SDXL-Turbo — kevrussell · 2026-10-27
- Diffusion LM ablation: sampler choice moves 25 points, noise only 0.8 — LucaAmb · 2026-10-27