{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:4ZAXKMRQEX3UBNINNER7WWVLJ3","short_pith_number":"pith:4ZAXKMRQ","schema_version":"1.0","canonical_sha256":"e64175323025f740b50d6923fb5aab4ee0584fd4493f7cdf02af0e3e2a6b087d","source":{"kind":"arxiv","id":"2605.15300","version":1},"attestation_state":"computed","paper":{"title":"Deep Pre-Alignment for VLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Deep Pre-Alignment replaces the ViT encoder with a small VLM perceiver to align visual features deeply with the LLM's text space.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Bo Zheng, Jun Song, Kaidong Zhang, Kechen Fang, Tianyu Yu, Yicheng Zhang, Yuan Yao, Zihao Wan","submitted_at":"2026-05-14T18:14:15Z","abstract_excerpt":"Most Vision Language Models (VLMs) directly map outputs from ViT encoders to the LLM via a lightweight projector. While effective, recent analysis suggests this architecture suffers from an alignment challenge: visual features remain distant from the text space in the initial layers of the LLM, forcing the model to waste critical depth~\\cite{zhang-etal-2024-investigating,artzy-schwartz-2024-attend} on superficial modality alignment rather than deep understanding and complex reasoning. In this work, we propose Deep Pre-Alignment (DPA), a novel architecture that replaces the standard ViT encoder"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.15300","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-14T18:14:15Z","cross_cats_sorted":[],"title_canon_sha256":"7e175ea771a9c586b737a13198eb361d0be7b9493deaa788087f9f5b5b77380d","abstract_canon_sha256":"017f88a96ca487ffa8bf0917da4183da73b1aac9cfb192fa146f4ce820b65e5e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:00:51.496929Z","signature_b64":"B5nHjq/XcyaAVpMgKlSxaBuoiQra+kHtJzP1/uCK4G2ETY1YoXiWcLFXuJJid40HMo8XMg7kmxJWznc4M9zBDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e64175323025f740b50d6923fb5aab4ee0584fd4493f7cdf02af0e3e2a6b087d","last_reissued_at":"2026-05-20T00:00:51.496070Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:00:51.496070Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Deep Pre-Alignment for VLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Deep Pre-Alignment replaces the ViT encoder with a small VLM perceiver to align visual features deeply with the LLM's text space.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Bo Zheng, Jun Song, Kaidong Zhang, Kechen Fang, Tianyu Yu, Yicheng Zhang, Yuan Yao, Zihao Wan","submitted_at":"2026-05-14T18:14:15Z","abstract_excerpt":"Most Vision Language Models (VLMs) directly map outputs from ViT encoders to the LLM via a lightweight projector. While effective, recent analysis suggests this architecture suffers from an alignment challenge: visual features remain distant from the text space in the initial layers of the LLM, forcing the model to waste critical depth~\\cite{zhang-etal-2024-investigating,artzy-schwartz-2024-attend} on superficial modality alignment rather than deep understanding and complex reasoning. In this work, we propose Deep Pre-Alignment (DPA), a novel architecture that replaces the standard ViT encoder"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On the 4B parameter scale, DPA outperforms baselines by 1.9 points across 8 multimodal benchmarks, with gains widening to 3.0 points at the 32B scale; by offloading alignment to the perceiver, DPA achieves a 32.9% reduction in language capability forgetting over 3 text benchmarks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That feeding the LLM with features from a small VLM perceiver (rather than a standard ViT plus projector) produces sufficiently deep pre-alignment so that the LLM's initial layers no longer perform superficial modality matching, as stated in the motivation citing prior alignment analyses.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Deep Pre-Alignment uses a small VLM perceiver instead of ViT to pre-align visual features with LLM text space, yielding 1.9-3.0 point gains on multimodal benchmarks and 32.9% less language forgetting.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Deep Pre-Alignment replaces the ViT encoder with a small VLM perceiver to align visual features deeply with the LLM's text space.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b77465daad9f712bb5b16d8fb34950afa971880d577f02f2d36ec11cbfa92020"},"source":{"id":"2605.15300","kind":"arxiv","version":1},"verdict":{"id":"43417420-73e3-4d11-bb77-c8d5f77883a7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T16:22:11.479040Z","strongest_claim":"On the 4B parameter scale, DPA outperforms baselines by 1.9 points across 8 multimodal benchmarks, with gains widening to 3.0 points at the 32B scale; by offloading alignment to the perceiver, DPA achieves a 32.9% reduction in language capability forgetting over 3 text benchmarks.","one_line_summary":"Deep Pre-Alignment uses a small VLM perceiver instead of ViT to pre-align visual features with LLM text space, yielding 1.9-3.0 point gains on multimodal benchmarks and 32.9% less language forgetting.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That feeding the LLM with features from a small VLM perceiver (rather than a standard ViT plus projector) produces sufficiently deep pre-alignment so that the LLM's initial layers no longer perform superficial modality matching, as stated in the motivation citing prior alignment analyses.","pith_extraction_headline":"Deep Pre-Alignment replaces the ViT encoder with a small VLM perceiver to align visual features deeply with the LLM's text space."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.15300/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T16:36:04.197308Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T16:31:18.354990Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T14:41:54.229768Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T13:33:22.780645Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"ad96a4f6fbe2a125ba7e09d379bc87aa1ff82d87e1bc43cc05c15a79938b428c"},"references":{"count":172,"sample":[{"doi":"","year":2023,"title":"International conference on machine learning , pages=","work_id":"99465f33-3e19-4fa9-8062-a629551b3315","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Flamingo: a visual language model for few-shot learning , author=. NeurIPS , volume=","work_id":"51b120de-c683-4609-8bae-7b6f4e485c79","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Changpinyo, Soravit and Sharma, Piyush and Ding, Nan and Soricut, Radu , booktitle=","work_id":"66d362ea-4f8a-4ef5-aceb-7bf05d542c72","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Byeon, Minwoo and Park, Beomhee and Kim, Haecheon and Lee, Sungjun and Baek, Woonhyuk and Kim, Saehoon , year =","work_id":"a7db0b62-446b-4fb5-b3d9-c23305eee65a","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Schuhmann, Christoph and Beaumont, Romain and Vencu, Richard and Gordon, Cade and Wightman, Ross and Cherti, Mehdi and Coombes, Theo and Katta, Aarush and Mullis, Clayton and Wortsman, Mitchell and ot","work_id":"c077101d-93a9-47bb-87af-d51c03cde237","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":172,"snapshot_sha256":"07a5382c0ba6ff713333afd313871102121e1770c29a22bcf2dc95c0bc5acc4e","internal_anchors":29},"formal_canon":{"evidence_count":2,"snapshot_sha256":"beb7c8506bdc7246677a083d000e68d4d48be25cf3c0b8111bff86024a81ae54"},"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":"2605.15300","created_at":"2026-05-20T00:00:51.496224+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.15300v1","created_at":"2026-05-20T00:00:51.496224+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.15300","created_at":"2026-05-20T00:00:51.496224+00:00"},{"alias_kind":"pith_short_12","alias_value":"4ZAXKMRQEX3U","created_at":"2026-05-20T00:00:51.496224+00:00"},{"alias_kind":"pith_short_16","alias_value":"4ZAXKMRQEX3UBNIN","created_at":"2026-05-20T00:00:51.496224+00:00"},{"alias_kind":"pith_short_8","alias_value":"4ZAXKMRQ","created_at":"2026-05-20T00:00:51.496224+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4ZAXKMRQEX3UBNINNER7WWVLJ3","json":"https://pith.science/pith/4ZAXKMRQEX3UBNINNER7WWVLJ3.json","graph_json":"https://pith.science/api/pith-number/4ZAXKMRQEX3UBNINNER7WWVLJ3/graph.json","events_json":"https://pith.science/api/pith-number/4ZAXKMRQEX3UBNINNER7WWVLJ3/events.json","paper":"https://pith.science/paper/4ZAXKMRQ"},"agent_actions":{"view_html":"https://pith.science/pith/4ZAXKMRQEX3UBNINNER7WWVLJ3","download_json":"https://pith.science/pith/4ZAXKMRQEX3UBNINNER7WWVLJ3.json","view_paper":"https://pith.science/paper/4ZAXKMRQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.15300&json=true","fetch_graph":"https://pith.science/api/pith-number/4ZAXKMRQEX3UBNINNER7WWVLJ3/graph.json","fetch_events":"https://pith.science/api/pith-number/4ZAXKMRQEX3UBNINNER7WWVLJ3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4ZAXKMRQEX3UBNINNER7WWVLJ3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4ZAXKMRQEX3UBNINNER7WWVLJ3/action/storage_attestation","attest_author":"https://pith.science/pith/4ZAXKMRQEX3UBNINNER7WWVLJ3/action/author_attestation","sign_citation":"https://pith.science/pith/4ZAXKMRQEX3UBNINNER7WWVLJ3/action/citation_signature","submit_replication":"https://pith.science/pith/4ZAXKMRQEX3UBNINNER7WWVLJ3/action/replication_record"}},"created_at":"2026-05-20T00:00:51.496224+00:00","updated_at":"2026-05-20T00:00:51.496224+00:00"}