{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:XKQHKK55KWAFKFJR23M2XUM7YW","short_pith_number":"pith:XKQHKK55","schema_version":"1.0","canonical_sha256":"baa0752bbd5580551531d6d9abd19fc593ec6510e2ed23dc614365d35a1f45d9","source":{"kind":"arxiv","id":"2504.01848","version":3},"attestation_state":"computed","paper":{"title":"PaperBench: Evaluating AI's Ability to Replicate AI Research","license":"http://creativecommons.org/licenses/by/4.0/","headline":"AI agents replicate only 21 percent of recent top AI research papers when starting from scratch.","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Amelia Glaese, Benjamin Kinsella, Dane Sherburn, Evan Mays, Giulio Starace, James Aung, Johannes Heidecke, Jun Shern Chan, Leon Maksin, Oliver Jaffe, Rachel Dias, Tejal Patwardhan, Wyatt Thompson","submitted_at":"2025-04-02T15:55:24Z","abstract_excerpt":"We introduce PaperBench, a benchmark evaluating the ability of AI agents to replicate state-of-the-art AI research. Agents must replicate 20 ICML 2024 Spotlight and Oral papers from scratch, including understanding paper contributions, developing a codebase, and successfully executing experiments. For objective evaluation, we develop rubrics that hierarchically decompose each replication task into smaller sub-tasks with clear grading criteria. In total, PaperBench contains 8,316 individually gradable tasks. Rubrics are co-developed with the author(s) of each ICML paper for accuracy and realism"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":true},"canonical_record":{"source":{"id":"2504.01848","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-04-02T15:55:24Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"dc269e44cb124817f631b9c1f198ba938035cc4594ee73b20aa8047ea4375577","abstract_canon_sha256":"b7772e59a981a43b3b67b065ee3e42aea25164a0b0a1fa578fcd63564266fb7f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:50.314929Z","signature_b64":"AgQZ6IYfiPeqXxneyYA9qdE8LyZ78j/RBqZwK2S+d2vnMwa9HbHiTE/AFeoD0xAX4fnkt80aHcBFJCES/zn5Bg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"baa0752bbd5580551531d6d9abd19fc593ec6510e2ed23dc614365d35a1f45d9","last_reissued_at":"2026-05-17T23:38:50.314464Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:50.314464Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"PaperBench: Evaluating AI's Ability to Replicate AI Research","license":"http://creativecommons.org/licenses/by/4.0/","headline":"AI agents replicate only 21 percent of recent top AI research papers when starting from scratch.","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Amelia Glaese, Benjamin Kinsella, Dane Sherburn, Evan Mays, Giulio Starace, James Aung, Johannes Heidecke, Jun Shern Chan, Leon Maksin, Oliver Jaffe, Rachel Dias, Tejal Patwardhan, Wyatt Thompson","submitted_at":"2025-04-02T15:55:24Z","abstract_excerpt":"We introduce PaperBench, a benchmark evaluating the ability of AI agents to replicate state-of-the-art AI research. Agents must replicate 20 ICML 2024 Spotlight and Oral papers from scratch, including understanding paper contributions, developing a codebase, and successfully executing experiments. For objective evaluation, we develop rubrics that hierarchically decompose each replication task into smaller sub-tasks with clear grading criteria. In total, PaperBench contains 8,316 individually gradable tasks. Rubrics are co-developed with the author(s) of each ICML paper for accuracy and realism"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the author-co-developed rubrics and the LLM judge together provide a reliable, unbiased measure of successful replication that generalizes beyond the 20 selected papers.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PaperBench is a new benchmark showing frontier AI agents replicate only 21% of tasks needed to reproduce state-of-the-art AI papers, below human expert performance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"AI agents replicate only 21 percent of recent top AI research papers when starting from scratch.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"071514f2404d401b85e9d6424f00a09fc9599133491b60f2e84d3924c62bfcbe"},"source":{"id":"2504.01848","kind":"arxiv","version":3},"verdict":{"id":"06b8b154-26d4-4c2b-aff4-c4e871f4e42e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T20:05:52.270470Z","strongest_claim":"the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0%.","one_line_summary":"PaperBench is a new benchmark showing frontier AI agents replicate only 21% of tasks needed to reproduce state-of-the-art AI papers, below human expert performance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the author-co-developed rubrics and the LLM judge together provide a reliable, unbiased measure of successful replication that generalizes beyond the 20 selected papers.","pith_extraction_headline":"AI agents replicate only 21 percent of recent top AI research papers when starting from scratch."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"6a03193263d61e6f953634c3d426c40e00f27d0d404963d19daaf61fef2d5872"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2504.01848","created_at":"2026-05-17T23:38:50.314538+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.01848v3","created_at":"2026-05-17T23:38:50.314538+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.01848","created_at":"2026-05-17T23:38:50.314538+00:00"},{"alias_kind":"pith_short_12","alias_value":"XKQHKK55KWAF","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"XKQHKK55KWAFKFJR","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"XKQHKK55","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":38,"internal_anchor_count":38,"sample":[{"citing_arxiv_id":"2605.17637","citing_title":"WebGameBench: Requirement-to-Application Evaluation for Coding Agents via Browser-Native Games","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2605.23204","citing_title":"AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery","ref_index":109,"is_internal_anchor":true},{"citing_arxiv_id":"2605.21413","citing_title":"Teaching AI Through Benchmark Construction: QuestBench as a Course-Based Practice for Accountable Knowledge Work","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2605.22343","citing_title":"Sibyl-AutoResearch: Autonomous Research Needs Self-Evolving Trial-and-Error Harnesses, Not Paper Generators","ref_index":36,"is_internal_anchor":true},{"citing_arxiv_id":"2605.21413","citing_title":"Teaching AI Through Benchmark Construction: QuestBench as a Course-Based Practice for Accountable Knowledge Work","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16616","citing_title":"MLReplicate: Benchmarking Autonomous Research Systems for Machine Learning Reproducibility","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2605.15871","citing_title":"Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design","ref_index":66,"is_internal_anchor":true},{"citing_arxiv_id":"2605.17637","citing_title":"WebGameBench: Requirement-to-Application Evaluation for Coding Agents via Browser-Native Games","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19156","citing_title":"How Far Are We From True Auto-Research?","ref_index":24,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16902","citing_title":"ArtifactLinker: Linking Scientific Artifacts for Automatic State-of-the-Art Discovery","ref_index":23,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02651","citing_title":"ARA: Agentic Reproducibility Assessment For Scalable Support Of Scientific Peer-Review","ref_index":47,"is_internal_anchor":true},{"citing_arxiv_id":"2506.22598","citing_title":"RExBench: Can coding agents autonomously implement AI research extensions?","ref_index":42,"is_internal_anchor":true},{"citing_arxiv_id":"2509.11206","citing_title":"Evalet: Evaluating Large Language Models through Functional Fragmentation","ref_index":76,"is_internal_anchor":true},{"citing_arxiv_id":"2509.20374","citing_title":"CFDLLMBench: A Benchmark Suite for Evaluating Large Language Models in Computational Fluid Dynamics","ref_index":49,"is_internal_anchor":true},{"citing_arxiv_id":"2510.24428","citing_title":"CodeWiki: Evaluating AI's Ability to Generate Holistic Documentation for Large-Scale Codebases","ref_index":44,"is_internal_anchor":true},{"citing_arxiv_id":"2602.00052","citing_title":"AI-assisted Protocol Information Extraction For Improved Accuracy and Efficiency in Clinical Trial Workflows","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2602.08561","citing_title":"Automating Computational Reproducibility in Social Science: Comparing Prompt-Based and Agent-Based Approaches","ref_index":29,"is_internal_anchor":true},{"citing_arxiv_id":"2602.11354","citing_title":"ReplicatorBench: Benchmarking LLM Agents for Replicability in Social and Behavioral Sciences","ref_index":15,"is_internal_anchor":true},{"citing_arxiv_id":"2602.21480","citing_title":"Both Ends Count! Just How Good are LLM Agents at \"Text-to-Big SQL\"?","ref_index":48,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13950","citing_title":"Collider-Bench: Benchmarking AI Agents with Particle Physics Analysis Reproduction","ref_index":24,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12808","citing_title":"Neurodata Without Boredom: Benchmarking Agentic AI for Data Reuse","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06607","citing_title":"AI CFD Scientist: Toward Open-Ended Computational Fluid Dynamics Discovery with Physics-Aware AI Agents","ref_index":30,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12808","citing_title":"Neurodata Without Boredom: Benchmarking Agentic AI for Data Reuse","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13275","citing_title":"ReproScore: Separating Readiness from Outcome in Research Software Reproducibility Assessment","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2604.00149","citing_title":"Towards Verifiable and Self-Correcting AI Physicists for Quantum Many-Body Simulations","ref_index":35,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/XKQHKK55KWAFKFJR23M2XUM7YW","json":"https://pith.science/pith/XKQHKK55KWAFKFJR23M2XUM7YW.json","graph_json":"https://pith.science/api/pith-number/XKQHKK55KWAFKFJR23M2XUM7YW/graph.json","events_json":"https://pith.science/api/pith-number/XKQHKK55KWAFKFJR23M2XUM7YW/events.json","paper":"https://pith.science/paper/XKQHKK55"},"agent_actions":{"view_html":"https://pith.science/pith/XKQHKK55KWAFKFJR23M2XUM7YW","download_json":"https://pith.science/pith/XKQHKK55KWAFKFJR23M2XUM7YW.json","view_paper":"https://pith.science/paper/XKQHKK55","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.01848&json=true","fetch_graph":"https://pith.science/api/pith-number/XKQHKK55KWAFKFJR23M2XUM7YW/graph.json","fetch_events":"https://pith.science/api/pith-number/XKQHKK55KWAFKFJR23M2XUM7YW/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XKQHKK55KWAFKFJR23M2XUM7YW/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XKQHKK55KWAFKFJR23M2XUM7YW/action/storage_attestation","attest_author":"https://pith.science/pith/XKQHKK55KWAFKFJR23M2XUM7YW/action/author_attestation","sign_citation":"https://pith.science/pith/XKQHKK55KWAFKFJR23M2XUM7YW/action/citation_signature","submit_replication":"https://pith.science/pith/XKQHKK55KWAFKFJR23M2XUM7YW/action/replication_record"}},"created_at":"2026-05-17T23:38:50.314538+00:00","updated_at":"2026-05-17T23:38:50.314538+00:00"}