{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.05162", "kind": "paper", "url": "https://arxiv.org/abs/2608.05162"}], "authors": ["Ayushi Agarwal"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.05162", "source_url": "https://arxiv.org/abs/2608.05162"}, "categories": ["General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "PoolBench isolates pooling strategies as the experimental variable in concept representation evaluation for decoder-only LLMs. It covers 17 concepts, 19 pooling strategies, and 3 open-weight models (Llama-3.1-8B, Gemma-2-9B, Mistral-7B) on a corpus of 37,693 real-text passages, with primary axis linear separability (AUROC) and diagnostic axes for steering and disentanglement."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.05162", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.05162", "title": "PoolBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.05162", "modality": null, "name": "PoolBench", "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": "unknown", "origin_record_id": "bm_poolbench_e79b8c50", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "70d30c3b37b3add19a13c95c6e3fca9dc683e8d4a2443c4e0840faeaa48b6772", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2608.05162"}, "publisher": null, "released": "2026-05-26", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.05162", "source_url": "https://arxiv.org/abs/2608.05162"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-05162", "source": "claire_radar", "source_benchmark_id": "2608.05162", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "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/2608.05162", "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"}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "b1f2e70efdb1ba23147bcbbf582544f71de38918836a8857adc217b0714068bb", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "PoolBench isolates pooling strategies as the experimental variable in concept representation evaluation for decoder-only LLMs. It covers 17 concepts, 19 pooling strategies, and 3 open-weight models (Llama-3.1-8B, Gemma-2-9B, Mistral-7B) on a corpus of 37,693 real-text passages, with primary axis linear separability (AUROC) and diagnostic axes for steering and disentanglement.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce PoolBench, a benchmark that isolates pooling as the experimental variable under a fixed evaluation protocol."}, "familyId": "bmf_9706d0afa65d", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_poolbench_e79b8c50", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2608.05162", "paper": "https://arxiv.org/abs/2608.05162", "pdf": "https://arxiv.org/pdf/2608.05162", "project": null, "report": "https://arxiv.org/abs/2608.05162"}, "metrics": [], "motivation": "Pooling is a consequential but under-examined design choice in decoder-only concept representation work: practitioners must collapse token-level hidden states into a passage-level vector, yet no shared protocol exists for comparing this choice across concepts, models, and tasks.", "name": "PoolBench", "oneLine": "PoolBench isolates pooling strategies as the experimental variable in concept representation evaluation for decoder-only LLMs. It covers 17 concepts, 19 pooling strategies, and 3 open-weight models (Llama-3.1-8B, Gemma-2-9B, Mistral-7B) on a corpus of 37,693 real-text passages, with primary axis linear separability (AUROC) and diagnostic axes for steering and disentanglement.", "paperTitle": "PoolBench: A Benchmark for Pooling Strategies in Concept Representation Evaluation for Decoder-Only LLMs", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.05162", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-05-26", "source": {"authors": ["Ayushi Agarwal"], "categories": ["cs.CL", "cs.LG"], "id": "2608.05162", "title": "PoolBench: A Benchmark for Pooling Strategies in Concept Representation Evaluation for Decoder-Only LLMs", "type": "arxiv", "url": "https://arxiv.org/abs/2608.05162"}, "sourceUpdatedAt": "2026-05-26T07:34:05Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "Pooling is a consequential but under-examined design choice in concept representation work, yet no shared protocol exists for comparing pooling rules across concepts, models, and tasks. PoolBench provides a controlled protocol with released corpus, pre-extracted activations, and scoring code to enable principled comparison."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
