{ "schema_version": 1, "builder_plan": { "scenario": "A self-hosted model uses more compute or memory than your workload can justify.", "change": "Check for a usable pruned checkpoint and runtime support before attempting model surgery.", "baseline": "The original unpruned model at the same precision, context, and load.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Task quality, actual GPU memory, serving throughput, and preparation cost.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2609.12310", "title": "ESTS at WMT26: Routing-Informed Expert Pruning for Model Compression", "summary": "The method replaces a full 20B-parameter mixture-of-experts model with a layer-wise pruned specialist, ranking experts by task-specific routing mass and allocating retained capacity via cross-lingual routing divergence, followed by recovery fine-tuning on synthetic data and MXFP4 quantization of the remaining expert weights. It costs a multi-stage pipeline: routing-statistics collection over task data, synthetic recovery-tuning data generated by a proprietary frontier model (GPT-5.1), retraining compute, and a custom inference stack with category inference, validation, retries, and fallback logic. It can fail when routing statistics do not transfer to unseen domains, when pruned experts carry rare-but-critical capability, when MXFP4 quantization compounds pruning error, or when the synthetic teacher data biases the specialist toward teacher artifacts, and the reported quality rests on pseudo-references from the same model family that generated the training data.", "rationale": "The pipeline depends on a specific MoE base model, proprietary-model-generated recovery data, and task-specific engineering effort, so it is worth tracking but not directly adoptable on a single 24 GB GPU without substantial replication work.", "technique": "Routing-informed MoE expert pruning with MXFP4 quantization", "family": "esparsidade_e_poda", "recommendation": "observar", "published": "2026-09-11", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.12310", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "The method replaces a full 20B-parameter mixture-of-experts model with a layer-wise pruned specialist, ranking experts by task-specific routing mass and allocating retained capacity via cross-lingual routing divergence, followed by recovery fine-tuning on synthetic data and MXFP4 quantization of the remaining expert weights. It costs a multi-stage pipeline: routing-statistics collection over task data, synthetic recovery-tuning data generated by a proprietary frontier model (GPT-5.1), retraining compute, and a custom inference stack with category inference, validation, retries, and fallback logic. It can fail when routing statistics do not transfer to unseen domains, when pruned experts carry rare-but-critical capability, when MXFP4 quantization compounds pruning error, or when the synthetic teacher data biases the specialist toward teacher artifacts, and the reported quality rests on pseudo-references from the same model family that generated the training data.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.12310", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Experts are physically removed from GPT-OSS-20B to reach 4.186B-7.770B total parameters (roughly 2.6x-4.8x parameter reduction) with packed artifacts of 4.55-6.33 GiB; quality is supported only by internal xCOMET-XL scoring against GPT-5.1 pseudo-references, not an independent benchmark.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.12310", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [ { "anchor": "S3.E1", "label": "(1)", "section": "3 Model Compression and Recovery", "role": "baseline", "latex": "\\bm{w}^{\\ell}_{i,\\varepsilon}=\\begin{cases}f(\\bm{h}^{\\ell}_{i})_{\\varepsilon}&\\text{if }\\varepsilon\\in\\text{top-}K\\\\\n\\ \\ \\ \\ 0&\\text{otherwise}\\end{cases}", "mathml": "𝒘i,εℓ={f​(𝒉iℓ)εif ​ε∈top-​K 0otherwise", "context": "In an MoE language model, at each layer ℓ∈{1,…,L} the router converts input hidden states 𝒉iℓ into logits 𝒛iℓ over all E experts. The top-K experts are then activated and aggregated using normalized weights. Accordingly, the weight of the output of each expert ε at token i can be defined as:", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "A1.E2", "label": "(2)", "section": "Appendix A Routing Divergence Metric", "role": "proposed_method", "latex": "\\text{Div}^{\\ell}_{t}=D_{\\mathrm{JS}}(\\bm{q}^{\\ell}_{t}||\\bm{q}^{\\ell}_{\\mathrm{Eng}})\\ \\ \\ \\ \\ \\in[0,1]", "mathml": "Divtℓ=DJS(𝒒tℓ||𝒒Engℓ)∈[0,1]", "context": "", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "A2.Ex2", "label": "", "section": "Appendix B Dynamic Capacity Allocation Details", "role": "proposed_method", "latex": "\\Delta c_{\\ell}=\\frac{d_{\\ell}^{s}}{\\sum_{j\\in\\mathcal{U}}d_{j}^{s}}B.", "mathml": "Δ​cℓ=dℓs∑j∈𝒰djs​B.", "context": "remaining expert slots to allocate. We then repeatedly distribute the remaining budget across layers that are not yet full. The upper bound for each cℓ is of course E, the number of experts in each MoE layer of the unpruned model. Let 𝒰={ℓ:cℓ<E} be the current set of non-full layers. For each ℓ∈𝒰, we allocate an additional real-valued capacity increment", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-09-29", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }