{
  "schema_version": "1.0",
  "id": "archive:https://arxiv.org/abs/2607.02513v1",
  "slug": "2607-02513v1-0lwaytn",
  "url": "https://feed7.dev/p/2607-02513v1-0lwaytn",
  "title": "LACUNA: A Testbed for Evaluating Localization Precision for LLM Unlearning",
  "why_included": "First unlearning testbed with ground-truth parameter localization: injects synthetic PII into known weights of OLMo 1B/7B models, showing current methods hide rather than erase and fall to resurfacing attacks.",
  "summary": "**The gist** LACUNA injects synthetic PII into predetermined parameters of **OLMo-based 1B and 7B** models via masked continual pretraining, giving unlearning research its **first ground-truth parameter-level** testbed instead of output-only benchmarks.",
  "practical_implication": "**Why it matters** If you count on unlearning to strip sensitive data from a model, output-level evals can lie: methods that look clean behaviorally were **highly imprecise** at the weight level and vulnerable to **resurfacing attacks**. When localization is right, even **simple gradient-based erasure** holds up — precision, not the removal algorithm, is the lever.",
  "agent_context": "**The gist** LACUNA injects synthetic PII into predetermined parameters of **OLMo-based 1B and 7B** models via masked continual pretraining, giving unlearning research its **first ground-truth parameter-level** testbed instead of output-only benchmarks.\n\n**Why it matters** If you count on unlearning to strip sensitive data from a model, output-level evals can lie: methods that look clean behaviorally were **highly imprecise** at the weight level and vulnerable to **resurfacing attacks**. When localization is right, even **simple gradient-based erasure** holds up — precision, not the removal algorithm, is the lever.\n\n**Watch out** The PII is **synthetic** and planted into known weights, a far cleaner setup than organically memorized data; whether precise localization is achievable in real pretrained models remains the open question.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2607.02513v1",
    "published_at": null
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "benchmark",
  "domains": [
    "research",
    "security"
  ],
  "topics": [
    "benchmark-integrity"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The PII is **synthetic** and planted into known weights, a far cleaner setup than organically memorized data; whether precise localization is achievable in real pretrained models remains the open question."
  ],
  "lifecycle": "Current",
  "published_at": null,
  "modified_at": null,
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2607-02513v1-0lwaytn",
    "json": "https://feed7.dev/p/2607-02513v1-0lwaytn.json",
    "markdown": "https://feed7.dev/p/2607-02513v1-0lwaytn.md"
  }
}