{"slug": "ests-at-wmt26-routing-informed-expert-pruning-for-model-compression", "title": "ESTS at WMT26: Routing-Informed Expert Pruning for Model Compression", "summary": "The ESTS team submitted six entries to the unconstrained WMT26 Model Compression Shared Task for English–Simplified Chinese and English–Egyptian Arabic, all derived from GPT-OSS-20B, with parameter counts ranging from 4.186B to 7.770B and packed artifact sizes from 4.55 to 6.33 GiB. The team ranked experts using task-specific routing mass and allocated retained capacity across layers via cross-lingual routing divergence before physically removing low-importance experts, then recovery-tuned the specialists on GPT-5.1-generated synthetic translation data and applied MXFP4 quantization to the retained expert projection weights. Internal xCOMET-XL evaluation using GPT-5.1 pseudo-references provided a comparison across the submitted compression operating points.", "body_md": "arXiv:2609.12310v1 Announce Type: new \nAbstract: We describe six submissions under the team name ESTS to the unconstrained WMT26 Model Compression Shared Task for English--Simplified Chinese and English--Egyptian Arabic. We submit three compression operating points per translation direction, all derived from GPT-OSS-20B. We use task-specific routing mass to rank experts and cross-lingual routing divergence to allocate retained capacity across layers, then physically remove low-importance experts. The resulting specialists are recovery-tuned on GPT-5.1-generated synthetic translation data and further compressed by applying MXFP4 quantization to the retained expert projection weights. We additionally implement a robust inference system for the instruction-conditioned WMT26 setting, including category inference, output validation, retries, segmented fallback, and source-owned JSON reconstruction. Across our six submissions, parameter counts range from 4.186B to 7.770B and packed artifact sizes from 4.55 to 6.33~GiB. Internal xCOMET-XL evaluation using GPT-5.1 pseudo-references provides an internal comparison across the submitted compression operating points.", "url": "https://wpnews.pro/news/ests-at-wmt26-routing-informed-expert-pruning-for-model-compression", "canonical_source": "https://arxiv.org/abs/2609.12310", "published_at": "2026-09-14 04:00:00+00:00", "updated_at": "2026-09-14 04:27:59.031933+00:00", "lang": "en", "topics": ["natural-language-processing", "large-language-models", "machine-learning", "ai-research"], "entities": ["ESTS", "WMT26 Model Compression Shared Task", "GPT-OSS-20B", "GPT-5.1", "xCOMET-XL"], "alternates": {"html": "https://wpnews.pro/news/ests-at-wmt26-routing-informed-expert-pruning-for-model-compression", "markdown": "https://wpnews.pro/news/ests-at-wmt26-routing-informed-expert-pruning-for-model-compression.md", "text": "https://wpnews.pro/news/ests-at-wmt26-routing-informed-expert-pruning-for-model-compression.txt", "jsonld": "https://wpnews.pro/news/ests-at-wmt26-routing-informed-expert-pruning-for-model-compression.jsonld"}}