{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.01161", "kind": "paper", "url": "https://arxiv.org/abs/2609.01161"}, {"id": "gh:spotai-mbzuai/copyshield", "kind": "repo", "url": "https://github.com/spotai-mbzuai/CopyShield.git"}], "authors": ["Maryam Alshehyari", "Dushyant Singh Chauhan", "Samuele Poppi", "Martin Takac", "Salem Lahlou", "Nils Lukas"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.01161", "source_url": "https://arxiv.org/abs/2609.01161"}, "categories": ["General AI", "Language & Knowledge", "cs.LG"], "description": {"en": "Compares copyright defenses in LLMs across three intervention levels using controlled memorization over five public-domain books and a shared protocol measuring leakage, utility, and degeneracy."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.01161", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.01161", "title": "CopyShield"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2609.01161", "modality": null, "name": "CopyShield", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "false", "origin_record_id": "bm_copyshield_e7fa34f9", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "f2cf1f8787894b0c0ccf5b2b64f2db376c46b35cd4cc7379a3b69b2c693f71f8", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-09-02T05:23:11.715633Z", "source_url": "https://arxiv.org/abs/2609.01161"}, "publisher": null, "released": "2026-09-01", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.01161", "source_url": "https://arxiv.org/abs/2609.01161"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-01161", "source": "claire_radar", "source_benchmark_id": "2609.01161", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/spotai-mbzuai/CopyShield", "githubScope": "benchmark_repo", "githubStars": 0, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2609.01161", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "Topical relevance to copyright issues in LLMs but unverified code availability reduces confidence in immediate attention.", "score": 50}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-02T05:23:11.715633Z", "inputHash": "530a51ee322ba3400ef964287bfa294b2a356eb8176fffb270c851f96ddf4ed2", "model": "deepseek-v4-pro"}, "curation": {"decisionReason": "The GitHub link status is unverified, and no data or independent release is provided; the benchmark cannot be confirmed as publicly reusable.", "model": "deepseek-v4-pro", "policy": "source-evidence-v2", "reviewedAt": "2026-09-02T05:23:11.715633Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Compares copyright defenses in LLMs across three intervention levels using controlled memorization over five public-domain books and a shared protocol measuring leakage, utility, and degeneracy.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce CopyShield, a controlled benchmark comparing three representative defenses at distinct intervention levels: contrastive decoding (output), Direct Preference Optimization (behavioral), and activation intervention (representation)."}, "familyId": "bmf_44bb3df7d69c", "firstSeenAt": "2026-09-02", "heat": null, "id": "bm_copyshield_e7fa34f9", "indexedAt": "2026-09-02T05:21:06.631990Z", "industrySectors": [], "links": {"code": "https://github.com/spotai-mbzuai/CopyShield.git", "data": null, "paper": "https://arxiv.org/abs/2609.01161", "pdf": "https://arxiv.org/pdf/2609.01161", "project": null, "report": "https://arxiv.org/abs/2609.01161"}, "metrics": [], "motivation": "Large language models can reproduce memorized text verbatim, yet copyright defenses are usually evaluated under incompatible protocols.", "name": "CopyShield", "oneLine": "Compares copyright defenses in LLMs across three intervention levels using controlled memorization over five public-domain books and a shared protocol measuring leakage, utility, and degeneracy.", "paperTitle": "CopyShield: A Cross-Level Benchmark of Copyright Defenses in LLMs", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.01161", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-02T05:21:06.631990Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.55, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.0, "value": 0}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 480, "score": 23}, "method": "attention-ranking-v13", "rank": 480, "score": 23}}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-09-01", "source": {"authors": ["Maryam Alshehyari", "Dushyant Singh Chauhan", "Samuele Poppi", "Martin Takac", "Salem Lahlou", "Nils Lukas"], "categories": ["cs.LG"], "id": "2609.01161", "title": "CopyShield: A Cross-Level Benchmark of Copyright Defenses in LLMs", "type": "arxiv", "url": "https://arxiv.org/abs/2609.01161"}, "sourceUpdatedAt": "2026-09-01T12:43:02Z", "topics": ["cs.LG"], "venueAttempts": [], "whyItMatters": "Provides cross-level reference baselines for copyright protection without sacrificing utility, highlighting targeted non-literal suppression as an open challenge."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
