{"record": {"aliases": ["Rushes: A Human Preference Dataset for Pluralistic Alignment"], "artifacts": [{"id": "arxiv:2607.20767", "kind": "paper", "url": "https://arxiv.org/abs/2607.20767"}, {"id": "gh:microsoft/rushes", "kind": "repo", "url": "https://github.com/microsoft/rushes"}], "categories": ["General AI", "Language & Knowledge", "Safety"], "description": {"en": "A dataset and diagnostic benchmark of 44,226 user decisions in AI-generated branching narratives. Models predict each user's choice at the event level."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.20767", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.20767", "title": "Rushes"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2607.20767", "modality": null, "name": "Rushes", "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_rushes-a-human-preference-dataset-for-plur_d3e08704", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "35ca3fdce1d87ef47081c0a7ea78ae3c65446aec206ec4eab6e57abf693d2e48", "relation": "evaluates_only", "review_model": "claude-code-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T20:45:51.600430Z", "source_url": "https://arxiv.org/abs/2607.20767"}, "publisher": null, "released": "2026-07-22", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.20767", "source_url": "https://arxiv.org/abs/2607.20767"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-20767", "source": "claire_radar", "source_benchmark_id": "2607.20767", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Rushes: A Human Preference Dataset for Pluralistic Alignment"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The source reports frontier LLMs losing to simple baselines on a pluralistic-alignment benchmark and gives a public code link.", "score": 55}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T20:45:51.600430Z", "inputHash": "700b3e47c89874b26a305fb1b22bdcf6980297683f808efe1401ae396f6a138b", "model": "claude-code-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "Rushes: A Human Preference Dataset for Pluralistic Alignment", "canonicalNameSource": "paper_title", "decisionReason": "The source names a benchmark, defines an event-level choice-prediction task with comparable percentage scores, and gives a public GitHub link. The release wording is future-tense, but the link is already supplied, which meets the rule.", "evaluationEvidence": {"reuse": "The full code for the platform and dataset will be available here: https://github.com/microsoft/rushes", "scoring": "frontier LLMs (34.23%) struggle to even match the Popularity Baseline (36.4%) on event-level choice prediction", "task": "select one choice from a small, explicit candidate set at each decision point"}, "model": "claude-code-default", "reviewedAt": "2026-09-24T20:45:51.600430Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A dataset and diagnostic benchmark of 44,226 user decisions in AI-generated branching narratives. Models predict each user's choice at the event level.", "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", "public artifact URL", "named entity is a model, method, framework, or dataset used in evaluation"], "snippet": "We introduce Rushes, a dataset and benchmark for studying revealed human engagement preferences in interactive narrative environments."}, "familyId": "bmf_8fb9979718ff", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_rushes-a-human-preference-dataset-for-plur_d3e08704", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/microsoft/rushes", "data": null, "paper": "https://arxiv.org/abs/2607.20767", "pdf": "https://arxiv.org/pdf/2607.20767", "project": null, "report": "https://arxiv.org/abs/2607.20767"}, "metrics": [], "motivation": "We introduce Rushes, a dataset and benchmark for studying revealed human engagement preferences in interactive narrative environments.", "name": "Rushes", "oneLine": "A dataset and diagnostic benchmark of 44,226 user decisions in AI-generated branching narratives. Models predict each user's choice at the event level.", "paperTitle": "Rushes: A Human Preference Dataset for Pluralistic Alignment", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.20767", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "publishers": [{"name": "Microsoft", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/microsoft/rushes"}], "readiness": "Runnable", "recognitionConfidence": 0.55, "relation": "evaluates_only", "releasedAt": "2026-07-22", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Michael Xu", "Jorge Leandro", "Sudha Rao", "Weijia Xu", "Nebojsa Jojic", "Gabriel DesGarennes", "Chris Quirk", "Bill Dolan"], "categories": ["cs.CL"], "id": "2607.20767", "title": "Rushes: A Human Preference Dataset for Pluralistic Alignment", "type": "arxiv", "url": "https://arxiv.org/abs/2607.20767"}, "sourceUpdatedAt": "2026-07-22T22:32:34Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "It tests whether models can predict individual, sequential engagement preferences instead of majority preferences. This makes it a probe for pluralistic alignment where reported LLMs trail simple baselines."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
