{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:ZAQMLWQTGYZJDZISVYUFWYRVRP","short_pith_number":"pith:ZAQMLWQT","canonical_record":{"source":{"id":"2604.22823","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-04-18T09:38:03Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"d2999f75856f3996a88ea668912c9163b48e49f04048958814acdd4313eb959b","abstract_canon_sha256":"3dbf9fb6300494f177deca23f0f23b321dfbb019f4bf28da0f1eb8b318b21148"},"schema_version":"1.0"},"canonical_sha256":"c820c5da13363291e512ae285b62358bd1ad53ac241a77359eb321a7e5554471","source":{"kind":"arxiv","id":"2604.22823","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2604.22823","created_at":"2026-07-14T02:21:34Z"},{"alias_kind":"arxiv_version","alias_value":"2604.22823v2","created_at":"2026-07-14T02:21:34Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.22823","created_at":"2026-07-14T02:21:34Z"},{"alias_kind":"pith_short_12","alias_value":"ZAQMLWQTGYZJ","created_at":"2026-07-14T02:21:34Z"},{"alias_kind":"pith_short_16","alias_value":"ZAQMLWQTGYZJDZIS","created_at":"2026-07-14T02:21:34Z"},{"alias_kind":"pith_short_8","alias_value":"ZAQMLWQT","created_at":"2026-07-14T02:21:34Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:ZAQMLWQTGYZJDZISVYUFWYRVRP","target":"record","payload":{"canonical_record":{"source":{"id":"2604.22823","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-04-18T09:38:03Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"d2999f75856f3996a88ea668912c9163b48e49f04048958814acdd4313eb959b","abstract_canon_sha256":"3dbf9fb6300494f177deca23f0f23b321dfbb019f4bf28da0f1eb8b318b21148"},"schema_version":"1.0"},"canonical_sha256":"c820c5da13363291e512ae285b62358bd1ad53ac241a77359eb321a7e5554471","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-14T02:21:34.588499Z","signature_b64":"7JTFHuF4mueB8sa9hZFEjpBASBQ5R1aXym95ddpi3m0ltxPxQcSPjiBy4V/mCBMK1rS8/9JfZxF1SPRF9hdWBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"c820c5da13363291e512ae285b62358bd1ad53ac241a77359eb321a7e5554471","last_reissued_at":"2026-07-14T02:21:34.587578Z","signature_status":"signed_v1","first_computed_at":"2026-07-14T02:21:34.587578Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2604.22823","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":"2026-07-14T02:21:34Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"CBQJF3c9TWBYd+/4iby0nmiKjxC6EJnWza2y0sgQPgfzUCVVPZnK43HB5g1zpX4RGuNmwwMNOzPHhdRGZMaQBg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-06T11:48:23.159259Z"},"content_sha256":"916d5a4aa1270e15c7c7372351a005284d4055d412d09d7e68533781b11137bf","schema_version":"1.0","event_id":"sha256:916d5a4aa1270e15c7c7372351a005284d4055d412d09d7e68533781b11137bf"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:ZAQMLWQTGYZJDZISVYUFWYRVRP","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"PivotMerge: Bridging Heterogeneous Multimodal Pre-training via Post-Alignment Model Merging","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"PivotMerge integrates cross-modal projectors from separate multimodal pre-training runs by separating shared alignment patterns from domain conflicts and weighting layers by their individual contributions.","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Baochen Xiong, Changsheng Xu, Haifeng Chen, Qimeng Zhang, Xiaoshan Yang, Yaguang Song, Zibo Shao","submitted_at":"2026-04-18T09:38:03Z","abstract_excerpt":"Multimodal Large Language Models (MLLMs) rely on multimodal pre-training over diverse data sources, where different datasets often induce complementary cross-modal alignment capabilities. Model merging provides a cost-effective mechanism for integrating multiple expert MLLMs with complementary strengths into a unified model. However, existing model merging research mainly focuses on post-finetuning scenarios, leaving the pre-training stage largely unexplored. We argue that the core of MLLM pre-training lies in establishing effective cross-modal alignment, which bridges visual and textual repre"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We propose PivotMerge, a post-alignment merging framework for cross-modal projectors that addresses cross-domain parameter interference and layer-wise alignment contribution disparity. Extensive experiments on multiple multimodal benchmarks show that PivotMerge consistently outperforms existing baselines.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the two identified challenges (cross-domain parameter interference and layer-wise alignment contribution disparity) are the primary obstacles in post-alignment merging and that the constructed CC12M-based evaluation scenarios sufficiently represent real heterogeneous multimodal pre-training distributions.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PivotMerge merges heterogeneous multimodal pre-trained models via shared-space decomposition to filter conflicts and layer-wise weights based on alignment contributions, outperforming baselines on multimodal benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"PivotMerge integrates cross-modal projectors from separate multimodal pre-training runs by separating shared alignment patterns from domain conflicts and weighting layers by their individual contributions.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"af59cac09510a775ea917c89e25510374025d3522a268799f99f823e19bdea83"},"source":{"id":"2604.22823","kind":"arxiv","version":2},"verdict":{"id":"c306614f-cf82-4dd7-a41d-f9ee9b39ad35","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T07:00:45.887250Z","strongest_claim":"We propose PivotMerge, a post-alignment merging framework for cross-modal projectors that addresses cross-domain parameter interference and layer-wise alignment contribution disparity. Extensive experiments on multiple multimodal benchmarks show that PivotMerge consistently outperforms existing baselines.","one_line_summary":"PivotMerge merges heterogeneous multimodal pre-trained models via shared-space decomposition to filter conflicts and layer-wise weights based on alignment contributions, outperforming baselines on multimodal benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the two identified challenges (cross-domain parameter interference and layer-wise alignment contribution disparity) are the primary obstacles in post-alignment merging and that the constructed CC12M-based evaluation scenarios sufficiently represent real heterogeneous multimodal pre-training distributions.","pith_extraction_headline":"PivotMerge integrates cross-modal projectors from separate multimodal pre-training runs by separating shared alignment patterns from domain conflicts and weighting layers by their individual contributions."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.22823/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"},"verdict_id":"c306614f-cf82-4dd7-a41d-f9ee9b39ad35"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-07-14T02:21:34Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"f7oWxlK39MRl01L/f4ZDUQZFGr5tgJ7WfJahZj78Srgqqu1ODIWNt4PWZlg9/tXdWDkOwaN2Tljxxj1g3FBFCw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-06T11:48:23.159906Z"},"content_sha256":"77d2079085e458d43eb11ad580ef5d808bf69a8f778bcae9107c22370a047436","schema_version":"1.0","event_id":"sha256:77d2079085e458d43eb11ad580ef5d808bf69a8f778bcae9107c22370a047436"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/ZAQMLWQTGYZJDZISVYUFWYRVRP/bundle.json","state_url":"https://pith.science/pith/ZAQMLWQTGYZJDZISVYUFWYRVRP/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/ZAQMLWQTGYZJDZISVYUFWYRVRP/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":"2026-08-06T11:48:23Z","links":{"resolver":"https://pith.science/pith/ZAQMLWQTGYZJDZISVYUFWYRVRP","bundle":"https://pith.science/pith/ZAQMLWQTGYZJDZISVYUFWYRVRP/bundle.json","state":"https://pith.science/pith/ZAQMLWQTGYZJDZISVYUFWYRVRP/state.json","well_known_bundle":"https://pith.science/.well-known/pith/ZAQMLWQTGYZJDZISVYUFWYRVRP/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:ZAQMLWQTGYZJDZISVYUFWYRVRP","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":"3dbf9fb6300494f177deca23f0f23b321dfbb019f4bf28da0f1eb8b318b21148","cross_cats_sorted":["cs.AI"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-04-18T09:38:03Z","title_canon_sha256":"d2999f75856f3996a88ea668912c9163b48e49f04048958814acdd4313eb959b"},"schema_version":"1.0","source":{"id":"2604.22823","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2604.22823","created_at":"2026-07-14T02:21:34Z"},{"alias_kind":"arxiv_version","alias_value":"2604.22823v2","created_at":"2026-07-14T02:21:34Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.22823","created_at":"2026-07-14T02:21:34Z"},{"alias_kind":"pith_short_12","alias_value":"ZAQMLWQTGYZJ","created_at":"2026-07-14T02:21:34Z"},{"alias_kind":"pith_short_16","alias_value":"ZAQMLWQTGYZJDZIS","created_at":"2026-07-14T02:21:34Z"},{"alias_kind":"pith_short_8","alias_value":"ZAQMLWQT","created_at":"2026-07-14T02:21:34Z"}],"graph_snapshots":[{"event_id":"sha256:77d2079085e458d43eb11ad580ef5d808bf69a8f778bcae9107c22370a047436","target":"graph","created_at":"2026-07-14T02:21:34Z","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":"We propose PivotMerge, a post-alignment merging framework for cross-modal projectors that addresses cross-domain parameter interference and layer-wise alignment contribution disparity. Extensive experiments on multiple multimodal benchmarks show that PivotMerge consistently outperforms existing baselines."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the two identified challenges (cross-domain parameter interference and layer-wise alignment contribution disparity) are the primary obstacles in post-alignment merging and that the constructed CC12M-based evaluation scenarios sufficiently represent real heterogeneous multimodal pre-training distributions."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"PivotMerge merges heterogeneous multimodal pre-trained models via shared-space decomposition to filter conflicts and layer-wise weights based on alignment contributions, outperforming baselines on multimodal benchmarks."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"PivotMerge integrates cross-modal projectors from separate multimodal pre-training runs by separating shared alignment patterns from domain conflicts and weighting layers by their individual contributions."}],"snapshot_sha256":"af59cac09510a775ea917c89e25510374025d3522a268799f99f823e19bdea83"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2604.22823/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Multimodal Large Language Models (MLLMs) rely on multimodal pre-training over diverse data sources, where different datasets often induce complementary cross-modal alignment capabilities. Model merging provides a cost-effective mechanism for integrating multiple expert MLLMs with complementary strengths into a unified model. However, existing model merging research mainly focuses on post-finetuning scenarios, leaving the pre-training stage largely unexplored. We argue that the core of MLLM pre-training lies in establishing effective cross-modal alignment, which bridges visual and textual repre","authors_text":"Baochen Xiong, Changsheng Xu, Haifeng Chen, Qimeng Zhang, Xiaoshan Yang, Yaguang Song, Zibo Shao","cross_cats":["cs.AI"],"headline":"PivotMerge integrates cross-modal projectors from separate multimodal pre-training runs by separating shared alignment patterns from domain conflicts and weighting layers by their individual contributions.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-04-18T09:38:03Z","title":"PivotMerge: Bridging Heterogeneous Multimodal Pre-training via Post-Alignment Model Merging"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2604.22823","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-10T07:00:45.887250Z","id":"c306614f-cf82-4dd7-a41d-f9ee9b39ad35","model_set":{"reader":"grok-4.3"},"one_line_summary":"PivotMerge merges heterogeneous multimodal pre-trained models via shared-space decomposition to filter conflicts and layer-wise weights based on alignment contributions, outperforming baselines on multimodal benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"PivotMerge integrates cross-modal projectors from separate multimodal pre-training runs by separating shared alignment patterns from domain conflicts and weighting layers by their individual contributions.","strongest_claim":"We propose PivotMerge, a post-alignment merging framework for cross-modal projectors that addresses cross-domain parameter interference and layer-wise alignment contribution disparity. Extensive experiments on multiple multimodal benchmarks show that PivotMerge consistently outperforms existing baselines.","weakest_assumption":"That the two identified challenges (cross-domain parameter interference and layer-wise alignment contribution disparity) are the primary obstacles in post-alignment merging and that the constructed CC12M-based evaluation scenarios sufficiently represent real heterogeneous multimodal pre-training distributions."}},"verdict_id":"c306614f-cf82-4dd7-a41d-f9ee9b39ad35"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:916d5a4aa1270e15c7c7372351a005284d4055d412d09d7e68533781b11137bf","target":"record","created_at":"2026-07-14T02:21:34Z","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":"3dbf9fb6300494f177deca23f0f23b321dfbb019f4bf28da0f1eb8b318b21148","cross_cats_sorted":["cs.AI"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-04-18T09:38:03Z","title_canon_sha256":"d2999f75856f3996a88ea668912c9163b48e49f04048958814acdd4313eb959b"},"schema_version":"1.0","source":{"id":"2604.22823","kind":"arxiv","version":2}},"canonical_sha256":"c820c5da13363291e512ae285b62358bd1ad53ac241a77359eb321a7e5554471","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"c820c5da13363291e512ae285b62358bd1ad53ac241a77359eb321a7e5554471","first_computed_at":"2026-07-14T02:21:34.587578Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-07-14T02:21:34.587578Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"7JTFHuF4mueB8sa9hZFEjpBASBQ5R1aXym95ddpi3m0ltxPxQcSPjiBy4V/mCBMK1rS8/9JfZxF1SPRF9hdWBw==","signature_status":"signed_v1","signed_at":"2026-07-14T02:21:34.588499Z","signed_message":"canonical_sha256_bytes"},"source_id":"2604.22823","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:916d5a4aa1270e15c7c7372351a005284d4055d412d09d7e68533781b11137bf","sha256:77d2079085e458d43eb11ad580ef5d808bf69a8f778bcae9107c22370a047436"],"state_sha256":"3148735405e0cae88edb11092aad6fb117c513a17d2a3b3a826b1ea642ba4106"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"Pj0NFnmH2ntvHjIiA4cz/xIBt/+4pmEReF9dt5iYaXUtHqkjzjyeeUkskKHZYletatgp9PKCcy6nboBehgjQBQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-08-06T11:48:23.165615Z","bundle_sha256":"5348fdc8831dc961d3f5722c62f1b1abdc85cb156c4f127f9188648d8f0f7d21"}}