{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:GPDP7VEYGGHKLNWOU6CRVUP22S","short_pith_number":"pith:GPDP7VEY","schema_version":"1.0","canonical_sha256":"33c6ffd498318ea5b6cea7851ad1fad4831e8bbe06e343e179bba4be984107cf","source":{"kind":"arxiv","id":"2503.07603","version":1},"attestation_state":"computed","paper":{"title":"Should VLMs be Pre-trained with Image Data?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Achal Dave, Benjamin Burchfiel, Igor Vasiljevic, Jean Mercat, Kushal Arora, Ludwig Schmidt, Russ Tedrake, Samir Yitzhak Gadre, Sedrick Keh, Shuran Song, Thomas Kollar","submitted_at":"2025-03-10T17:58:19Z","abstract_excerpt":"Pre-trained LLMs that are further trained with image data perform well on vision-language tasks. While adding images during a second training phase effectively unlocks this capability, it is unclear how much of a gain or loss this two-step pipeline gives over VLMs which integrate images earlier into the training process. To investigate this, we train models spanning various datasets, scales, image-text ratios, and amount of pre-training done before introducing vision tokens. We then fine-tune these models and evaluate their downstream performance on a suite of vision-language and text-only tas"},"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":false},"canonical_record":{"source":{"id":"2503.07603","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2025-03-10T17:58:19Z","cross_cats_sorted":[],"title_canon_sha256":"cf97b1de6a8c57600894242db7fef59fae3d07eb8722b11ca38d1d62ab9fd9df","abstract_canon_sha256":"7f453f96e03703b277274f9447e5d7f42b21c4048ba305525030ff9adc0cc61c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:28:07.932860Z","signature_b64":"BE6xxmhME84micYHGrb02/p3UzgpBA15Y/ogZpEpw6alVjGNPG4WkKxC0wK8rv/K9LOvpbHNT8D+eUYlPoVcBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"33c6ffd498318ea5b6cea7851ad1fad4831e8bbe06e343e179bba4be984107cf","last_reissued_at":"2026-07-05T10:28:07.931860Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:28:07.931860Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Should VLMs be Pre-trained with Image Data?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Achal Dave, Benjamin Burchfiel, Igor Vasiljevic, Jean Mercat, Kushal Arora, Ludwig Schmidt, Russ Tedrake, Samir Yitzhak Gadre, Sedrick Keh, Shuran Song, Thomas Kollar","submitted_at":"2025-03-10T17:58:19Z","abstract_excerpt":"Pre-trained LLMs that are further trained with image data perform well on vision-language tasks. While adding images during a second training phase effectively unlocks this capability, it is unclear how much of a gain or loss this two-step pipeline gives over VLMs which integrate images earlier into the training process. To investigate this, we train models spanning various datasets, scales, image-text ratios, and amount of pre-training done before introducing vision tokens. We then fine-tune these models and evaluate their downstream performance on a suite of vision-language and text-only tas"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.07603","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2503.07603/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"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":"2503.07603","created_at":"2026-07-05T10:28:07.932030+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.07603v1","created_at":"2026-07-05T10:28:07.932030+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.07603","created_at":"2026-07-05T10:28:07.932030+00:00"},{"alias_kind":"pith_short_12","alias_value":"GPDP7VEYGGHK","created_at":"2026-07-05T10:28:07.932030+00:00"},{"alias_kind":"pith_short_16","alias_value":"GPDP7VEYGGHKLNWO","created_at":"2026-07-05T10:28:07.932030+00:00"},{"alias_kind":"pith_short_8","alias_value":"GPDP7VEY","created_at":"2026-07-05T10:28:07.932030+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2604.19728","citing_title":"VLA Foundry: A Unified Framework for Training Vision-Language-Action Models","ref_index":30,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/GPDP7VEYGGHKLNWOU6CRVUP22S","json":"https://pith.science/pith/GPDP7VEYGGHKLNWOU6CRVUP22S.json","graph_json":"https://pith.science/api/pith-number/GPDP7VEYGGHKLNWOU6CRVUP22S/graph.json","events_json":"https://pith.science/api/pith-number/GPDP7VEYGGHKLNWOU6CRVUP22S/events.json","paper":"https://pith.science/paper/GPDP7VEY"},"agent_actions":{"view_html":"https://pith.science/pith/GPDP7VEYGGHKLNWOU6CRVUP22S","download_json":"https://pith.science/pith/GPDP7VEYGGHKLNWOU6CRVUP22S.json","view_paper":"https://pith.science/paper/GPDP7VEY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.07603&json=true","fetch_graph":"https://pith.science/api/pith-number/GPDP7VEYGGHKLNWOU6CRVUP22S/graph.json","fetch_events":"https://pith.science/api/pith-number/GPDP7VEYGGHKLNWOU6CRVUP22S/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/GPDP7VEYGGHKLNWOU6CRVUP22S/action/timestamp_anchor","attest_storage":"https://pith.science/pith/GPDP7VEYGGHKLNWOU6CRVUP22S/action/storage_attestation","attest_author":"https://pith.science/pith/GPDP7VEYGGHKLNWOU6CRVUP22S/action/author_attestation","sign_citation":"https://pith.science/pith/GPDP7VEYGGHKLNWOU6CRVUP22S/action/citation_signature","submit_replication":"https://pith.science/pith/GPDP7VEYGGHKLNWOU6CRVUP22S/action/replication_record"}},"created_at":"2026-07-05T10:28:07.932030+00:00","updated_at":"2026-07-05T10:28:07.932030+00:00"}