Value-and-Structure Alignment for Routing-Consistent Quantization of Mixture-of-Experts Models

MoE quantization error is amplified discontinuously: a tiny perturbation can flip the top-k expert choice, changing the entire computation path for that token.

A PTQ OBJECTIVE (not an architecture change) that penalizes divergence of post-quantization expert selection from pre-quantization selection, in two complementary terms: VALUE alignment matches the routing-relevant logits/scores themselves, and STRUCTURE alignment preserves expert ORDERING and the top-k decision boundaries. Unlike methods that quantize routing scores directly, VSRAQ adds no inference-time overhead and plugs into existing quantization frameworks.

Paper's own claim (abstract): 'unlike dense models, MoE models are sensitive to routing instability', and preserving routing consistency 'reduces quantization-induced degradation without introducing any inference-time overhead'. No numeric accuracy or compression figure in the abstract, so speedup_reported stays null. The zero-inference-overhead property is what separates it from EAQuant's routing-consistency penalty, which alters the objective during calibration but is also inference-free - the real difference is the loss form.

arXiv ID 2606.05688 verified via /abs/. Venue: Semantic Scholar returns arXiv.org only - arxiv-preprint, year = v1 year 2026. Affiliation not verified. arXiv ID and title verified by fetching https://arxiv.org/abs/2606.05688 and reading the citation_title meta tag on 2026-10-03. Venue and year from the Semantic Scholar graph API batch endpoint (paper/arXiv:2606.05688) unless the record's notes say otherwise. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.