{ "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": "2610.19121", "title": "Unlocking Structured Pruning for Reasoning Language Models via Attention-Preserving MLP Trimming", "summary": "LayerTrim replaces training-based structured pruning pipelines (e.g., RESP) with a training-free method that trims MLP features in redundant decoder layers while preserving attention blocks and depth, yielding a smaller dense model. It costs implementation effort for the layer-selection and trimming procedure and incurs some accuracy loss relative to the dense model, though less than competing pruning methods; no recovery training or extra GPU memory is required. It can fail on reasoning workloads or model families not covered in the evaluation, since long-form reasoning accuracy (e.g., AIME) is sensitive to layer removal and the reported gains come from a single study on three models.", "rationale": "A training-free pruning method that runs on a single GPU and preserves reasoning accuracy fits constrained infrastructure, but the claimed speedup is modest and task-dependent, so the reader should validate it on their own model and workloads before production adoption.", "technique": "Attention-preserving structured layer MLP pruning", "family": "esparsidade_e_poda", "recommendation": "testar", "published": "2026-10-13", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.19121", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "LayerTrim replaces training-based structured pruning pipelines (e.g., RESP) with a training-free method that trims MLP features in redundant decoder layers while preserving attention blocks and depth, yielding a smaller dense model. It costs implementation effort for the layer-selection and trimming procedure and incurs some accuracy loss relative to the dense model, though less than competing pruning methods; no recovery training or extra GPU memory is required. It can fail on reasoning workloads or model families not covered in the evaluation, since long-form reasoning accuracy (e.g., AIME) is sensitive to layer removal and the reported gains come from a single study on three models.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.19121", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "1.24x faster prefill and 19.7% lower peak GPU memory on Qwen3-14B versus the dense model, with average accuracy up to 21.4 points above the RESP pruning baseline.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.19121", "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": "" }