{ "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": "2610.18516", "title": "Two For The Price Of One: Curvature-Aware Pruning Preserves Influence Functions Too", "summary": "CACTUS replaces standard post-training pruning solvers (and the LLM-Surgeon optimizer) with an ADMM-based method that jointly matches layer activations and backpropagated gradients, supporting both pruning and sparse-plus-low-rank decompositions. It costs a calibration optimization pass per layer, added implementation complexity over magnitude or WANDA-style pruning, and optionally memory for the low-rank correction term. It can fail if the calibration data is unrepresentative of deployment inputs, if the target hardware lacks sparse-kernel support to realize 2:4 speedups, or if influence-function preservation does not transfer to tasks beyond the evaluated data-selection and perplexity settings.", "rationale": "It fits single-GPU post-training workflows and shows measurable quality gains over existing pruning at 50% sparsity, but adoption should follow workload-specific validation of calibration cost and sparse-inference support.", "technique": "Curvature-aware pruning via ADMM (CACTUS)", "family": "esparsidade_e_poda", "recommendation": "testar", "published": "2026-10-13", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.18516", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "CACTUS replaces standard post-training pruning solvers (and the LLM-Surgeon optimizer) with an ADMM-based method that jointly matches layer activations and backpropagated gradients, supporting both pruning and sparse-plus-low-rank decompositions. It costs a calibration optimization pass per layer, added implementation complexity over magnitude or WANDA-style pruning, and optionally memory for the low-rank correction term. It can fail if the calibration data is unrepresentative of deployment inputs, if the target hardware lacks sparse-kernel support to realize 2:4 speedups, or if influence-function preservation does not transfer to tasks beyond the evaluated data-selection and perplexity settings.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.18516", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Runs over 28x faster than LLM-Surgeon on substantially weaker hardware, while reducing the C4 perplexity gap to the dense model by 45% at 50% pruning of Qwen3-4B.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.18516", "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": "" }