{ "schema_version": 1, "builder_plan": { "scenario": "You already need task-specific training and want to reduce its memory or compute cost.", "change": "Change only the training technique on a small controlled run after checking prerequisites.", "baseline": "Your current training recipe with the same model, data split, and evaluation.", "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": "Held-out quality, peak memory, training time, and full compute cost to the target quality.", "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": "2610.26370", "title": "An Empirical Study of Parameter-Efficient Fine-Tuning for Mixture-of-Experts LLMs", "summary": "The recipe replaces full fine-tuning of MoE models (and adapters that bypass experts) with LoRA/DoRA applied directly to native expert weights, trained without a load-balancing auxiliary loss since no nonzero coefficient Pareto-dominated the zero baseline. Cost is limited to a small trainable-parameter budget (17% fewer than the module-bypass baseline) within a single-GPU fine-tuning run, plus a measured 3.2-point average downstream regression for the redesigned variant. Failure modes include expert routing drift degrading general-knowledge retention, results transferring poorly across the three studied backbones to other MoE architectures, and task-specific sensitivity to which experts and ranks are adapted.", "rationale": "The reader can run LoRA/DoRA on expert weights of an open MoE model on a 24 GB GPU, but the optimal module granularity, rank, and retention trade-off depend on their specific backbone and downstream tasks, so small-scale validation is required before production adoption.", "technique": "PEFT on MoE experts via native expert-weight LoRA/DoRA", "family": "treino_eficiente", "recommendation": "testar", "published": "2026-10-19", "as_of": "2026-10-20", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.26370", "independent_implementations": 0, "total_implementations": 1, "stars": 0, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "The recipe replaces full fine-tuning of MoE models (and adapters that bypass experts) with LoRA/DoRA applied directly to native expert weights, trained without a load-balancing auxiliary loss since no nonzero coefficient Pareto-dominated the zero baseline. Cost is limited to a small trainable-parameter budget (17% fewer than the module-bypass baseline) within a single-GPU fine-tuning run, plus a measured 3.2-point average downstream regression for the redesigned variant. Failure modes include expert routing drift degrading general-knowledge retention, results transferring poorly across the three studied backbones to other MoE architectures, and task-specific sensitivity to which experts and ranks are adapted.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.26370", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Adapting MoE experts alone yields the highest downstream performance with competitive general-capability retention; a weight-level redesign of MoE^2-LoRA improves four upstream endpoints by 2.4 to 17.9 points with 17% fewer trainable parameters, at a 3.2-point average downstream cost.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.26370", "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": [], "equations_status": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }