{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.24232", "kind": "paper", "url": "https://arxiv.org/abs/2608.24232"}], "authors": ["Zhenyu Wu", "Siyuan Chen", "Changchun Yang", "Jiaqi Dong", "Min Zhou", "Ali Almadan", "Talal Hammad", "Faisal Wahbo", "Aminullah Tora", "Mona Alshahrani", "Xin Gao"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.24232", "source_url": "https://arxiv.org/abs/2608.24232"}, "categories": ["General AI", "Reasoning", "Safety", "Safety & Trustworthiness"], "description": {"en": "TRACE is an evidence-grounded safety evaluation benchmark covering prompts, reasoning traces, and final responses of Large Reasoning Models (LRMs). It includes prompts in two languages, nine risk categories, ten attack strategies, and per-component safety annotations with supporting evidence. Evaluates 18 guardrail models on safety judgment and evidence extraction."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.24232", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.24232", "title": "TRACE"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.24232", "modality": null, "name": "TRACE", "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_trace_e2618c65", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "8f87e6eb3222006d10849d62f39d9e52140db7519a162da7813207a326ddaa1b", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-08-27T04:18:56.318436Z", "source_url": "https://arxiv.org/abs/2608.24232"}, "publisher": null, "released": "2026-08-25", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.24232", "source_url": "https://arxiv.org/abs/2608.24232"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-24232", "source": "claire_radar", "source_benchmark_id": "2608.24232", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 2, "hfPaperUrl": "https://huggingface.co/papers/2608.24232", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-09-25", "state": "stale"}}, "status": "stale"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Safety-critical and timely topic with detailed benchmark design; likely to interest safety and alignment researchers.", "score": 65}, "capabilities": ["Reasoning"], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-27T04:22:25.189955Z", "inputHash": "334c57270ff5753aad73450c83179a58448f704276a05508875567d44b2f8a82", "model": "deepseek-v4-flash"}, "curation": {"canonicalNameEvidence": "we introduce TRACE, an evidence-grounded safety evaluation benchmark that covers the entire LRM inference pipeline: prompts, reasoning traces, and final responses.", "canonicalNameSource": "abstract", "decisionReason": "The benchmark has a clearly defined evaluation object, stable scoring contract, and explicit release statement, though no code or data link is provided in the artifact evidence.", "model": "deepseek-v4-pro", "reviewedAt": "2026-08-27T04:18:56.318436Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "TRACE is an evidence-grounded safety evaluation benchmark covering prompts, reasoning traces, and final responses of Large Reasoning Models (LRMs). It includes prompts in two languages, nine risk categories, ten attack strategies, and per-component safety annotations with supporting evidence. Evaluates 18 guardrail models on safety judgment and evidence extraction.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "To address these limitations, we introduce TRACE, an evidence-grounded safety evaluation benchmark that covers the entire LRM inference pipeline: prompts, reasoning traces, and final responses."}, "familyId": "bmf_eafe895eb811", "firstSeenAt": "2026-08-26", "heat": null, "id": "bm_trace_e2618c65", "indexedAt": "2026-08-26T06:08:22.244942Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2608.24232", "pdf": "https://arxiv.org/pdf/2608.24232", "project": null, "report": "https://arxiv.org/abs/2608.24232"}, "metrics": [], "motivation": "Large Reasoning Models (LRMs) generate intermediate reasoning traces that may contain unsafe content, even when their final responses appear safe.", "name": "TRACE", "oneLine": "TRACE is an evidence-grounded safety evaluation benchmark covering prompts, reasoning traces, and final responses of Large Reasoning Models (LRMs). It includes prompts in two languages, nine risk categories, ten attack strategies, and per-component safety annotations with supporting evidence. Evaluates 18 guardrail models on safety judgment and evidence extraction.", "paperTitle": "TRACE: An Evidence-Grounded Benchmark for Safety Evaluation of Large Reasoning Models", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.24232", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-26T06:08:22.244942Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.15, "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": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.19411886932144953, "value": 2}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 45}, "method": "attention-ranking-v13", "rank": null, "score": 45}}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-25", "source": {"authors": ["Zhenyu Wu", "Siyuan Chen", "Changchun Yang", "Jiaqi Dong", "Min Zhou", "Ali Almadan", "Talal Hammad", "Faisal Wahbo", "Aminullah Tora", "Mona Alshahrani", "Xin Gao"], "categories": ["cs.AI"], "id": "2608.24232", "title": "TRACE: An Evidence-Grounded Benchmark for Safety Evaluation of Large Reasoning Models", "type": "arxiv", "url": "https://arxiv.org/abs/2608.24232"}, "sourceUpdatedAt": "2026-08-25T08:35:47Z", "topics": ["Safety", "Reasoning"], "venueAttempts": [], "whyItMatters": "Safety evaluation of LRMs currently focuses on prompts and final responses, ignoring reasoning traces that may contain unsafe content. TRACE provides a reproducible benchmark with evidence annotations, enabling precise evaluation of guardrails and highlighting challenges in trace-level safety judgment."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
