{"record": {"aliases": ["Instruction Stacking Collapse: A Benchmark and the Capability-Dependent Value of Prompt Compilation"], "artifacts": [{"id": "arxiv:2608.02639", "kind": "paper", "url": "https://arxiv.org/abs/2608.02639"}], "authors": ["Atul Anand", "Sourav Chattaraj"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.02639", "source_url": "https://arxiv.org/abs/2608.02639"}, "categories": ["Code & Software", "Software & AI Compute", "cs.SE"], "description": {"en": "A repeatable evaluation of instruction following as multiple verifier-checked constraints accumulate in a prompt."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.02639", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.02639", "title": "Instruction Stacking Collapse"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.02639", "modality": null, "name": "Instruction Stacking Collapse", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_instruction-stacking-collapse-a-benchmark-_ace8079d", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "c881dfe9ad39c438b78d1e6541143e8488613dcf57f944aaa8d865ae6a2a06f3", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T19:26:48.696948Z", "source_url": "https://arxiv.org/abs/2608.02639"}, "publisher": null, "released": "2026-07-31", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.02639", "source_url": "https://arxiv.org/abs/2608.02639"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-02639", "source": "claire_radar", "source_benchmark_id": "2608.02639", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Instruction Stacking Collapse: A Benchmark and the Capability-Dependent Value of Prompt Compilation"], "annotation": "Unknown", "applicationDomains": ["Software & AI Compute"], "area": "Code & Software", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The evaluation addresses a practical production prompting failure mode across several widely used model families, but near-term attention cannot be established from the supplied source alone.", "score": 6}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T19:26:02.417471Z", "inputHash": "aec73142cf92ea02bf3fdd69033625ace57681785c6079c253a6c0b7940a4912", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "Instruction Stacking Collapse", "canonicalNameSource": "paper_title", "decisionReason": "The paper formally names Instruction Stacking Collapse in its title, specifies verifier-checked tasks and a comparable follow-rate measure, and explicitly states that the benchmark and supporting artifacts are released for full reproduction.", "evaluationEvidence": {"reuse": "We release the benchmark, verifiers, and cached runs for full reproduction.", "scoring": "the follow rate falls from ~96% to as low as 20%", "task": "stacks 24 verifier-checked instructions, one to twenty at a time"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T19:26:48.696948Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A repeatable evaluation of instruction following as multiple verifier-checked constraints accumulate in a prompt.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce a benchmark that stacks 24 verifier-checked instructions, one to twenty at a time, and evaluate three production-tier LLMs (Claude Sonnet 4.6, GPT-5-mini, Gemini 2.5 Flash)."}, "familyId": "bmf_b92913ec6b1b", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_instruction-stacking-collapse-a-benchmark-_ace8079d", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Software & Cloud"], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2608.02639", "pdf": "https://arxiv.org/pdf/2608.02639", "project": null, "report": "https://arxiv.org/abs/2608.02639"}, "metrics": [], "motivation": "Production prompts rarely carry a single instruction.", "name": "Instruction Stacking Collapse", "oneLine": "A repeatable evaluation of instruction following as multiple verifier-checked constraints accumulate in a prompt.", "paperTitle": "Instruction Stacking Collapse: A Benchmark and the Capability-Dependent Value of Prompt Compilation", "primaryDomain": "Software & AI Compute", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.02639", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-07-31", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Atul Anand", "Sourav Chattaraj"], "categories": ["cs.SE", "cs.AI"], "id": "2608.02639", "title": "Instruction Stacking Collapse: A Benchmark and the Capability-Dependent Value of Prompt Compilation", "type": "arxiv", "url": "https://arxiv.org/abs/2608.02639"}, "sourceUpdatedAt": "2026-07-31T10:58:29Z", "topics": ["cs.SE"], "venueAttempts": [], "whyItMatters": "It measures nonlinear degradation and reproducible conflicts among simultaneous prompt constraints, while supporting full reproduction with released verifiers and cached runs."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
