MoEQuant: Enhancing Quantization for Mixture-of-Experts Large Language Models via Expert-Balanced Sampling and Affinity Guidance

PTQ degrades MoE models far worse than dense models, and the usual dense calibration set starves rarely-activated experts.

Names two MoE-specific calibration pathologies and fixes both: (1) INTER-EXPERT IMBALANCE - token routing is uneven, so a random calibration set under-samples and biases rarely-used experts, fixed by Expert-Balanced Self-Sampling (EBSS) which constructs a calibration set whose per-expert sample counts match expected routing frequency; (2) INTRA-EXPERT IMBALANCE - the weighted aggregation of a token's experts makes the effective per-expert signal differ from the raw activations, fixed by Affinity Guidance which scores calibration samples by how strongly they align with their assigned expert.

Paper's own claim (abstract): PTQ on MoE suffers severe accuracy degradation and diminished generalization, and MoEQuant is a quantization framework tailored for MoE LLMs addressing both imbalances. The abstract gives NO numeric accuracy delta or throughput figure, so speedup_reported stays descriptive. The transferable contribution is the diagnosis: MoE PTQ fails on calibration coverage, not on weight bit width.

arXiv ID 2505.03804 verified via /abs/. Venue ICML 2025 from Semantic Scholar; year = venue year 2025 (v1 2025-05-02). Repo already has source qlab2-arxiv-moequant. arXiv ID and title verified by fetching https://arxiv.org/abs/2505.03804 and reading the citation_title meta tag on 2026-10-03. Venue and year from the Semantic Scholar graph API batch endpoint (paper/arXiv:2505.03804) unless the record's notes say otherwise. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: icml has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.