{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:U4ZQRXTAQTMTDQE4WS66F2FVVP","short_pith_number":"pith:U4ZQRXTA","canonical_record":{"source":{"id":"2603.07433","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-08T03:10:39Z","cross_cats_sorted":["cs.CV"],"title_canon_sha256":"0a17e150362ea75f1b651945cd110cac508b18037eaff452b8329cd41f0fb872","abstract_canon_sha256":"1c86de336f20c56f2fb81bfefa1043abefce8647b7745cc9d990d5953f398ca3"},"schema_version":"1.0"},"canonical_sha256":"a73308de6084d931c09cb4bde2e8b5abd154fcf377459d97b3887ca245db10d2","source":{"kind":"arxiv","id":"2603.07433","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2603.07433","created_at":"2026-05-18T03:09:22Z"},{"alias_kind":"arxiv_version","alias_value":"2603.07433v2","created_at":"2026-05-18T03:09:22Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.07433","created_at":"2026-05-18T03:09:22Z"},{"alias_kind":"pith_short_12","alias_value":"U4ZQRXTAQTMT","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"U4ZQRXTAQTMTDQE4","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"U4ZQRXTA","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:U4ZQRXTAQTMTDQE4WS66F2FVVP","target":"record","payload":{"canonical_record":{"source":{"id":"2603.07433","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-08T03:10:39Z","cross_cats_sorted":["cs.CV"],"title_canon_sha256":"0a17e150362ea75f1b651945cd110cac508b18037eaff452b8329cd41f0fb872","abstract_canon_sha256":"1c86de336f20c56f2fb81bfefa1043abefce8647b7745cc9d990d5953f398ca3"},"schema_version":"1.0"},"canonical_sha256":"a73308de6084d931c09cb4bde2e8b5abd154fcf377459d97b3887ca245db10d2","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:09:22.959183Z","signature_b64":"WEfzsBTEJWsBMOFewOO0rm3HswImHgWTVMUuWY1ehP6HaVbpOHPMsFSGTKWXboxtE4i0btB921j0m/o8RGg3Dg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a73308de6084d931c09cb4bde2e8b5abd154fcf377459d97b3887ca245db10d2","last_reissued_at":"2026-05-18T03:09:22.958437Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:09:22.958437Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2603.07433","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T03:09:22Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"7/h9ipPPXnk0vDvssdl/Nbv96QCb5GTZDR/Snn6XHfA+LkbJJdnHLhflUKsBIjyByghZz621AojqLcOI66QgDQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-26T03:15:36.287505Z"},"content_sha256":"575e7010650fb6cc9c88a257e9e413cbbc073cd92647010e3ae3a01ae8c86dea","schema_version":"1.0","event_id":"sha256:575e7010650fb6cc9c88a257e9e413cbbc073cd92647010e3ae3a01ae8c86dea"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:U4ZQRXTAQTMTDQE4WS66F2FVVP","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Data Agent: Learning to Select Data via End-to-End Dynamic Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Data Agent learns to select training samples dynamically as a sequential decision problem guided by evolving loss and uncertainty rewards.","cross_cats":["cs.CV"],"primary_cat":"cs.LG","authors_text":"Baile Xu, Fangjian Su, Furao Shen, Hai Gan, Jie Li, Soujanya Poria, Suorong Yang, Ziqi Ye","submitted_at":"2026-03-08T03:10:39Z","abstract_excerpt":"Dynamic Data selection aims to accelerate training by prioritizing informative samples during online training. However, existing methods typically rely on task-specific handcrafted metrics or static/snapshot-based criteria to estimate sample importance, limiting scalability across learning paradigms and making it difficult to capture the evolving utility of data throughout training. To address this challenge, we propose Data Agent, an end-to-end dynamic data selection framework that formulates data selection as a training-aware sequential decision-making problem. The agent learns a sample-wise"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Data Agent consistently accelerates training while preserving or improving performance, e.g., reducing costs by over 50% on ImageNet-1k and MMLU with lossless performance.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a composite reward combining loss-based difficulty and confidence-based uncertainty, together with a tuning-free adaptive weighting mechanism, can reliably capture the evolving utility of each sample throughout training across diverse tasks and architectures.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Data Agent learns a co-evolving sample selection policy end-to-end that accelerates training by over 50% on ImageNet-1k and MMLU with no performance loss.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Data Agent learns to select training samples dynamically as a sequential decision problem guided by evolving loss and uncertainty rewards.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"38c565c458142f8fb6c0162a59445ced4224056d5b6be82fe2112da4336ed515"},"source":{"id":"2603.07433","kind":"arxiv","version":2},"verdict":{"id":"9f37a904-185c-4db8-98ad-93be03931e5a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T15:20:01.761598Z","strongest_claim":"Data Agent consistently accelerates training while preserving or improving performance, e.g., reducing costs by over 50% on ImageNet-1k and MMLU with lossless performance.","one_line_summary":"Data Agent learns a co-evolving sample selection policy end-to-end that accelerates training by over 50% on ImageNet-1k and MMLU with no performance loss.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a composite reward combining loss-based difficulty and confidence-based uncertainty, together with a tuning-free adaptive weighting mechanism, can reliably capture the evolving utility of each sample throughout training across diverse tasks and architectures.","pith_extraction_headline":"Data Agent learns to select training samples dynamically as a sequential decision problem guided by evolving loss and uncertainty rewards."},"references":{"count":19,"sample":[{"doi":"","year":null,"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","ref_index":1,"cited_arxiv_id":"2303.08774","is_internal_anchor":true},{"doi":"","year":null,"title":"Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks","work_id":"b77fede6-b151-43c2-8d1e-2aef2cdd28f8","ref_index":2,"cited_arxiv_id":"1711.02257","is_internal_anchor":true},{"doi":"","year":2010,"title":"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale","work_id":"e96730e3-129b-4db6-b981-15ab7932e297","ref_index":3,"cited_arxiv_id":"2010.11929","is_internal_anchor":true},{"doi":"","year":null,"title":"Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators","work_id":"ef25adcf-addb-445e-b3b5-858eeb9883ca","ref_index":4,"cited_arxiv_id":"2404.04475","is_internal_anchor":true},{"doi":"","year":2009,"title":"Measuring Massive Multitask Language Understanding","work_id":"e87ec49a-544b-4ec8-8991-75298c64ff5e","ref_index":5,"cited_arxiv_id":"2009.03300","is_internal_anchor":true}],"resolved_work":19,"snapshot_sha256":"0806e3532f680562e788b25f99115c200d23b9696294315bc0103a2938d4b901","internal_anchors":8},"formal_canon":{"evidence_count":2,"snapshot_sha256":"7945a3ee50f0e893c41293fb497087924d38aa9cf5f544e9e64ac413178f88c7"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"9f37a904-185c-4db8-98ad-93be03931e5a"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T03:09:22Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"290itFpLjuGSoMbjITlRxpAVYFiz5cyrnyKU65ULgastcxBjgxVGtT0wKgy3zLEJnI2OFGvYPHHoYab1KtwRCw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-26T03:15:36.288388Z"},"content_sha256":"ccbaf7f1bc11192e5b634845a731fcfb85e462ffcd97a5c6acf25ab98cfea92b","schema_version":"1.0","event_id":"sha256:ccbaf7f1bc11192e5b634845a731fcfb85e462ffcd97a5c6acf25ab98cfea92b"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/U4ZQRXTAQTMTDQE4WS66F2FVVP/bundle.json","state_url":"https://pith.science/pith/U4ZQRXTAQTMTDQE4WS66F2FVVP/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/U4ZQRXTAQTMTDQE4WS66F2FVVP/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-05-26T03:15:36Z","links":{"resolver":"https://pith.science/pith/U4ZQRXTAQTMTDQE4WS66F2FVVP","bundle":"https://pith.science/pith/U4ZQRXTAQTMTDQE4WS66F2FVVP/bundle.json","state":"https://pith.science/pith/U4ZQRXTAQTMTDQE4WS66F2FVVP/state.json","well_known_bundle":"https://pith.science/.well-known/pith/U4ZQRXTAQTMTDQE4WS66F2FVVP/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:U4ZQRXTAQTMTDQE4WS66F2FVVP","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"1c86de336f20c56f2fb81bfefa1043abefce8647b7745cc9d990d5953f398ca3","cross_cats_sorted":["cs.CV"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-08T03:10:39Z","title_canon_sha256":"0a17e150362ea75f1b651945cd110cac508b18037eaff452b8329cd41f0fb872"},"schema_version":"1.0","source":{"id":"2603.07433","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2603.07433","created_at":"2026-05-18T03:09:22Z"},{"alias_kind":"arxiv_version","alias_value":"2603.07433v2","created_at":"2026-05-18T03:09:22Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.07433","created_at":"2026-05-18T03:09:22Z"},{"alias_kind":"pith_short_12","alias_value":"U4ZQRXTAQTMT","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"U4ZQRXTAQTMTDQE4","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"U4ZQRXTA","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:ccbaf7f1bc11192e5b634845a731fcfb85e462ffcd97a5c6acf25ab98cfea92b","target":"graph","created_at":"2026-05-18T03:09:22Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"Data Agent consistently accelerates training while preserving or improving performance, e.g., reducing costs by over 50% on ImageNet-1k and MMLU with lossless performance."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That a composite reward combining loss-based difficulty and confidence-based uncertainty, together with a tuning-free adaptive weighting mechanism, can reliably capture the evolving utility of each sample throughout training across diverse tasks and architectures."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Data Agent learns a co-evolving sample selection policy end-to-end that accelerates training by over 50% on ImageNet-1k and MMLU with no performance loss."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Data Agent learns to select training samples dynamically as a sequential decision problem guided by evolving loss and uncertainty rewards."}],"snapshot_sha256":"38c565c458142f8fb6c0162a59445ced4224056d5b6be82fe2112da4336ed515"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"7945a3ee50f0e893c41293fb497087924d38aa9cf5f544e9e64ac413178f88c7"},"paper":{"abstract_excerpt":"Dynamic Data selection aims to accelerate training by prioritizing informative samples during online training. However, existing methods typically rely on task-specific handcrafted metrics or static/snapshot-based criteria to estimate sample importance, limiting scalability across learning paradigms and making it difficult to capture the evolving utility of data throughout training. To address this challenge, we propose Data Agent, an end-to-end dynamic data selection framework that formulates data selection as a training-aware sequential decision-making problem. The agent learns a sample-wise","authors_text":"Baile Xu, Fangjian Su, Furao Shen, Hai Gan, Jie Li, Soujanya Poria, Suorong Yang, Ziqi Ye","cross_cats":["cs.CV"],"headline":"Data Agent learns to select training samples dynamically as a sequential decision problem guided by evolving loss and uncertainty rewards.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-08T03:10:39Z","title":"Data Agent: Learning to Select Data via End-to-End Dynamic Optimization"},"references":{"count":19,"internal_anchors":8,"resolved_work":19,"sample":[{"cited_arxiv_id":"2303.08774","doi":"","is_internal_anchor":true,"ref_index":1,"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","year":null},{"cited_arxiv_id":"1711.02257","doi":"","is_internal_anchor":true,"ref_index":2,"title":"Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks","work_id":"b77fede6-b151-43c2-8d1e-2aef2cdd28f8","year":null},{"cited_arxiv_id":"2010.11929","doi":"","is_internal_anchor":true,"ref_index":3,"title":"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale","work_id":"e96730e3-129b-4db6-b981-15ab7932e297","year":2010},{"cited_arxiv_id":"2404.04475","doi":"","is_internal_anchor":true,"ref_index":4,"title":"Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators","work_id":"ef25adcf-addb-445e-b3b5-858eeb9883ca","year":null},{"cited_arxiv_id":"2009.03300","doi":"","is_internal_anchor":true,"ref_index":5,"title":"Measuring Massive Multitask Language Understanding","work_id":"e87ec49a-544b-4ec8-8991-75298c64ff5e","year":2009}],"snapshot_sha256":"0806e3532f680562e788b25f99115c200d23b9696294315bc0103a2938d4b901"},"source":{"id":"2603.07433","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-15T15:20:01.761598Z","id":"9f37a904-185c-4db8-98ad-93be03931e5a","model_set":{"reader":"grok-4.3"},"one_line_summary":"Data Agent learns a co-evolving sample selection policy end-to-end that accelerates training by over 50% on ImageNet-1k and MMLU with no performance loss.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Data Agent learns to select training samples dynamically as a sequential decision problem guided by evolving loss and uncertainty rewards.","strongest_claim":"Data Agent consistently accelerates training while preserving or improving performance, e.g., reducing costs by over 50% on ImageNet-1k and MMLU with lossless performance.","weakest_assumption":"That a composite reward combining loss-based difficulty and confidence-based uncertainty, together with a tuning-free adaptive weighting mechanism, can reliably capture the evolving utility of each sample throughout training across diverse tasks and architectures."}},"verdict_id":"9f37a904-185c-4db8-98ad-93be03931e5a"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:575e7010650fb6cc9c88a257e9e413cbbc073cd92647010e3ae3a01ae8c86dea","target":"record","created_at":"2026-05-18T03:09:22Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"1c86de336f20c56f2fb81bfefa1043abefce8647b7745cc9d990d5953f398ca3","cross_cats_sorted":["cs.CV"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-08T03:10:39Z","title_canon_sha256":"0a17e150362ea75f1b651945cd110cac508b18037eaff452b8329cd41f0fb872"},"schema_version":"1.0","source":{"id":"2603.07433","kind":"arxiv","version":2}},"canonical_sha256":"a73308de6084d931c09cb4bde2e8b5abd154fcf377459d97b3887ca245db10d2","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"a73308de6084d931c09cb4bde2e8b5abd154fcf377459d97b3887ca245db10d2","first_computed_at":"2026-05-18T03:09:22.958437Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T03:09:22.958437Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"WEfzsBTEJWsBMOFewOO0rm3HswImHgWTVMUuWY1ehP6HaVbpOHPMsFSGTKWXboxtE4i0btB921j0m/o8RGg3Dg==","signature_status":"signed_v1","signed_at":"2026-05-18T03:09:22.959183Z","signed_message":"canonical_sha256_bytes"},"source_id":"2603.07433","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:575e7010650fb6cc9c88a257e9e413cbbc073cd92647010e3ae3a01ae8c86dea","sha256:ccbaf7f1bc11192e5b634845a731fcfb85e462ffcd97a5c6acf25ab98cfea92b"],"state_sha256":"c77323b4463f8eeda9ad245200dbb4b40eef546526977c4eb3e997a8a37bd344"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"CI3vqKFCk4t9a6EsP1O0mHqoEe0LerPNnbTXEoKP68Tb0Q3YkbxjCqljIEDG+RbuOaQFXyDuex2+nXFxtqw6Cg==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-26T03:15:36.292428Z","bundle_sha256":"10371fd28ed4a15124f47e349706ae3f2dfe22f6dd99b0c0bb5d9f9d23748701"}}