{"record": {"aliases": ["Reconstructing the Right Episode: Evaluating Interleaved Conversational Memory Beyond Long Context"], "artifacts": [{"id": "arxiv:2608.25655", "kind": "paper", "url": "https://arxiv.org/abs/2608.25655"}], "authors": ["Zhexi Feng", "Ruiyi Zhang", "Yongbo Yang", "Pengtao Xie"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.25655", "source_url": "https://arxiv.org/abs/2608.25655"}, "categories": ["General AI", "Language & Knowledge", "Long Context"], "description": {"en": "SCALE-QA evaluates interleaved conversational memory using 3,000 audited multiple-choice questions across 10 domains, where correct answers depend on causally related evidence from earlier turns in flat unsegmented threads."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.25655", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.25655", "title": "SCALE-QA"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.25655", "modality": null, "name": "SCALE-QA", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_reconstructing-the-right-episode-evaluatin_98f96c99", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "326fc39eddd32c7d01779889c97f4750646da4cc88e8caea39a494a7b35a78e5", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-08-27T04:12:36.511123Z", "source_url": "https://arxiv.org/abs/2608.25655"}, "publisher": null, "released": "2026-08-26", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.25655", "source_url": "https://arxiv.org/abs/2608.25655"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-25655", "source": "claire_radar", "source_benchmark_id": "2608.25655", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Reconstructing the Right Episode: Evaluating Interleaved Conversational Memory Beyond Long Context"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "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": null, "hfPaperUrl": "https://huggingface.co/papers/2608.25655", "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", "state": "unavailable"}}, "status": "unavailable"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The arXiv paper describes a concrete benchmark with 3,000 questions, deterministic scoring, and strong baseline comparisons across multiple LLMs, which is likely to attract moderate attention among NLP researchers working on long-context memory.", "score": 65}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-27T04:12:36.511123Z", "inputHash": "1805af50c05d20ee7e6fe8b74902055190078f6a89a885240da7ce01a950d7fd", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "We introduce SCALE-QA, a constraint-grounded task QA benchmark for flat unsegmented threads targeting episode integrity failure.", "canonicalNameSource": "abstract", "decisionReason": "The benchmark introduces a named dataset with a clear evaluation object, deterministic four-way multiple-choice grading, and an explicit release of questions and runtime builder, enabling reuse by other teams.", "model": "deepseek-v4-pro", "reviewedAt": "2026-08-27T04:12:36.511123Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SCALE-QA evaluates interleaved conversational memory using 3,000 audited multiple-choice questions across 10 domains, where correct answers depend on causally related evidence from earlier turns in flat unsegmented threads.", "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 SCALE-QA, a constraint-grounded task QA benchmark for flat unsegmented threads targeting episode integrity failure."}, "familyId": "bmf_c199dc017445", "firstSeenAt": "2026-08-27", "heat": null, "id": "bm_reconstructing-the-right-episode-evaluatin_98f96c99", "indexedAt": "2026-08-27T04:12:10.575570Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2608.25655", "pdf": "https://arxiv.org/pdf/2608.25655", "project": null, "report": "https://arxiv.org/abs/2608.25655"}, "metrics": [], "motivation": "Conversations with chat assistants increasingly span many topics in a single long-running thread, challenging memory systems.", "name": "SCALE-QA", "oneLine": "SCALE-QA evaluates interleaved conversational memory using 3,000 audited multiple-choice questions across 10 domains, where correct answers depend on causally related evidence from earlier turns in flat unsegmented threads.", "paperTitle": "Reconstructing the Right Episode: Evaluating Interleaved Conversational Memory Beyond Long Context", "primaryDomain": "General AI", "publication": {"evidence": "19 pages, 6 figures, 30 tables. Accepted to the Main Conference of EMNLP 2026", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2608.25655", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "Main Conference of EMNLP 2026", "verifiedAt": "2026-08-27T04:12:10.575570Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.0, "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": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-08-26", "source": {"authors": ["Zhexi Feng", "Ruiyi Zhang", "Yongbo Yang", "Pengtao Xie"], "categories": ["cs.CL", "cs.AI"], "id": "2608.25655", "title": "Reconstructing the Right Episode: Evaluating Interleaved Conversational Memory Beyond Long Context", "type": "arxiv", "url": "https://arxiv.org/abs/2608.25655"}, "sourceUpdatedAt": "2026-08-26T11:37:03Z", "topics": ["Long Context"], "venueAttempts": [{"decisionRaw": "19 pages, 6 figures, 30 tables. Accepted to the Main Conference of EMNLP 2026", "evidence": [{"level": "author-claim", "observedAt": "2026-08-27T04:12:10.575570Z", "rawValue": "19 pages, 6 figures, 30 tables. Accepted to the Main Conference of EMNLP 2026", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2608.25655"}], "reviewStatus": "accepted", "venueName": "Main Conference of EMNLP 2026"}], "whyItMatters": "It tests episode integrity failure in long, mixed-topic conversations, a harder memory regime than benchmarks with explicit topic boundaries, and provides deterministic grading for comparing QA performance."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
