{ "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": "2610.28773", "title": "Building a Nested LADDER for LLM Unstructured Post-Training Pruning", "summary": "It replaces uniform or heuristic layer-wise sparsity allocation in post-training pruning (e.g., standard SparseGPT allocation) with a greedy procedure that, at each step, commits the layer update causing the least end-to-end degradation per removed parameter, producing exactly nested masks across budgets. The cost is added pruning-time computation—repeated candidate evaluations across layers at every sparsity step—plus integration complexity on top of a backend such as SparseGPT; inference cost of the pruned model is unchanged. It can fail at benign sparsity (10-60%), where it is comparable or slightly weaker than baselines, and the high-sparsity benefit did not survive distillation on one of five models.", "rationale": "The method is implementable on a single 24 GB GPU for small-to-mid models and is relevant if the reader needs high-sparsity CPU-friendly deployment, but the advantage only materializes at 70%+ sparsity and requires workload-specific validation, so it should be tested rather than adopted outright.", "technique": "Damage-driven incremental layer-wise sparsity allocation", "family": "esparsidade_e_poda", "recommendation": "testar", "published": "2026-10-20", "as_of": "2026-10-24", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.28773", "independent_implementations": 0, "total_implementations": 1, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "It replaces uniform or heuristic layer-wise sparsity allocation in post-training pruning (e.g., standard SparseGPT allocation) with a greedy procedure that, at each step, commits the layer update causing the least end-to-end degradation per removed parameter, producing exactly nested masks across budgets. The cost is added pruning-time computation—repeated candidate evaluations across layers at every sparsity step—plus integration complexity on top of a backend such as SparseGPT; inference cost of the pruned model is unchanged. It can fail at benign sparsity (10-60%), where it is comparable or slightly weaker than baselines, and the high-sparsity benefit did not survive distillation on one of five models.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.28773", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "At high sparsity (70% onward), LADDER is substantially more robust than uniform and adaptive allocation baselines, and this advantage largely persists after posterior distillation across four of five models, with improved retention of aggregate zero-shot accuracy; no numeric factor is stated in the abstract.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.28773", "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": "" }