{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.31022", "kind": "paper", "url": "https://arxiv.org/abs/2608.31022"}], "authors": ["Vernon Toh", "Navonil Majumder", "Zhengyuan Liu", "Nancy F. Chen", "Soujanya Poria"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.31022", "source_url": "https://arxiv.org/abs/2608.31022"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "Evaluates agents' perceptual state construction in partially observable environments through a sequential glimpse-based MNIST digit search with lookback constraints, scored on recognition accuracy across memory representations."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.31022", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.31022", "title": "MNIST-PRO"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.31022", "modality": null, "name": "MNIST-PRO", "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": "false", "origin_record_id": "bm_mnist-pro_997cca4c", "origin_source": "arxiv", "recognition_confidence": "0.85", "record_sha256": "9ab4e9d6f7784b19ae423019c55d21744d4b902cb80e6880095235468a25a79b", "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/2608.31022"}, "publisher": null, "released": "2026-08-31", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.31022", "source_url": "https://arxiv.org/abs/2608.31022"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-31022", "source": "claire_radar", "source_benchmark_id": "2608.31022", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-09-02", "attemptedAt": "2026-09-02T05:12:51.621571Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": "2026-09-01T00:00:00.000Z", "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 7, "hfPaperUrl": "https://huggingface.co/papers/2608.31022", "observedAt": "2026-09-02T05:12:51.621571Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-09-02T05:12:51.621571Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "Leveraging the familiar MNIST dataset could attract researchers focused on agent perception and memory, though the benchmark's novelty and niche scope may limit initial spread.", "score": 60}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-02T05:23:11.715633Z", "inputHash": "40ee779f7e7a85c911115324293a223729da8875cf874e0f1276f9c995e67c88", "model": "deepseek-v4-pro"}, "curation": {"decisionReason": "The benchmark is explicitly defined with a clear task protocol and evaluation criteria, despite lacking artifact links in the input; its formal description in the paper provides a reusable protocol.", "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": "Evaluates agents' perceptual state construction in partially observable environments through a sequential glimpse-based MNIST digit search with lookback constraints, scored on recognition accuracy across memory representations.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "evaluation protocol evidence"], "snippet": "We address this with MNIST-PRO, a benchmark that isolates agentic perception by converting MNIST digit recognition into a sequential, glimpse-based search task with lookback constraints."}, "familyId": "bmf_18e28f0a46a8", "firstSeenAt": "2026-09-01", "heat": null, "id": "bm_mnist-pro_997cca4c", "indexedAt": "2026-09-02T05:21:06.631990Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2608.31022", "pdf": "https://arxiv.org/pdf/2608.31022", "project": null, "report": "https://arxiv.org/abs/2608.31022"}, "metrics": [], "motivation": "AI agents in partially observable environments need to coordinate active sensing with working memory to maintain an evolving perceptual state.", "name": "MNIST-PRO", "oneLine": "Evaluates agents' perceptual state construction in partially observable environments through a sequential glimpse-based MNIST digit search with lookback constraints, scored on recognition accuracy across memory representations.", "paperTitle": "MNIST-PRO: MNIST is Back as a Partially Observable World for AI Agents", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.31022", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-02T05:21:06.631990Z"}, "publications": [], "ranking": {"30d": {"confidence": "Low", "coverage": 0.3, "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": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.36857068284431604, "value": 7}}, "confidence": "Low", "coverage": 0.3, "rank": null, "score": 46}, "method": "attention-ranking-v13", "rank": null, "score": 46}, "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": 0.36857068284431604, "value": 7}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 48}, "method": "attention-ranking-v13", "rank": null, "score": 48}}, "readiness": "Paper only", "recognitionConfidence": 0.85, "relation": "introduces", "releasedAt": "2026-08-31", "source": {"authors": ["Vernon Toh", "Navonil Majumder", "Zhengyuan Liu", "Nancy F. Chen", "Soujanya Poria"], "categories": ["cs.AI", "cs.CV"], "id": "2608.31022", "title": "MNIST-PRO: MNIST is Back as a Partially Observable World for AI Agents", "type": "arxiv", "url": "https://arxiv.org/abs/2608.31022"}, "sourceUpdatedAt": "2026-08-31T16:05:39Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "Isolates perceptual-state construction from physical and control complexity, exposing fundamental bottlenecks in active sensing, memory integration, and belief revision that are hidden in full-observability benchmarks."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
