{ "schema_version": 1, "builder_plan": { "scenario": "Start with one repeated product task that has a measurable failure or cost.", "change": "Read the methods and identify one change you can isolate in that workflow.", "baseline": "Your current implementation on the same held-out tasks.", "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": "Successful outcomes, response time, human rework, and total cost per successful task.", "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": "2609.38895", "title": "Unmerge: Efficient Machine Unlearning via Task Arithmetic", "summary": "Unmerge replaces gradient-based approximate unlearning (with its data-dependent hyperparameter search) and full retraining by subtracting a low-rank factorized forget task vector from the merged finetuned model, optimizing closed-form objectives that bound forget leakage and retain damage. It costs a per-layer low-rank basis computation plus a lightweight optimization, runs faster than stronger unlearning baselines, and avoids retraining; the residual is a small tail-eigenvalue error that can perturb retain activations. It can fail when forget and retain knowledge are so entangled that no low-rank separation exists, and evidence for LLMs is limited to scaling on one 3B model rather than validated unlearning quality.", "rationale": "The method runs on a single consumer GPU and scales to Llama-3.2-3B, but its value depends entirely on whether the reader has a regulatory or product need for post-hoc unlearning, and its guarantees are demonstrated mainly on image classifiers.", "technique": "Low-rank task-arithmetic machine unlearning (Unmerge)", "family": "outro", "recommendation": "testar", "published": "2026-09-30", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.38895", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "Unmerge replaces gradient-based approximate unlearning (with its data-dependent hyperparameter search) and full retraining by subtracting a low-rank factorized forget task vector from the merged finetuned model, optimizing closed-form objectives that bound forget leakage and retain damage. It costs a per-layer low-rank basis computation plus a lightweight optimization, runs faster than stronger unlearning baselines, and avoids retraining; the residual is a small tail-eigenvalue error that can perturb retain activations. It can fail when forget and retain knowledge are so entangled that no low-rank separation exists, and evidence for LLMs is limited to scaling on one 3B model rather than validated unlearning quality.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.38895", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "On class-level unlearning (ResNet-50, CIFAR-100 and Tiny ImageNet), Unmerge improves the Tug-of-War metric by up to ~24% over a comparable-runtime baseline and up to ~18% over baselines running ~5x slower, with membership-inference exposure at retraining level.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.38895", "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": "" }