{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:J2A5EGLFRJRKYDAACPYE6ABCTM","short_pith_number":"pith:J2A5EGLF","schema_version":"1.0","canonical_sha256":"4e81d219658a62ac0c0013f04f00229b27cb8e84219c055bab85f053938dad0f","source":{"kind":"arxiv","id":"2605.12374","version":2},"attestation_state":"computed","paper":{"title":"Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Granular alignment at three levels lets MLLMs generate stable visual latents by fixing decoder-to-input mismatch.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CV","authors_text":"Dexin Wang, Guanjun Jiang, Hao Li, Lei Lv, Li Wang, Mengyu Zhou, Pascal Poupart, Qi Zhao, Xiaoxi Jiang, Yanting Miao, Yutao Sun","submitted_at":"2026-05-12T16:41:09Z","abstract_excerpt":"Visual latent reasoning lets a multimodal large language model (MLLM) create intermediate visual evidence as continuous tokens, avoiding external tools or image generators. However, existing methods usually follow an output-as-input latent paradigm and yield unstable gains. We identify evidence for a feature-space mismatch that can contribute to this instability: dominant visual-latent models build on pre-norm MLLMs and reuse decoder hidden states as predicted latent inputs, even though these states occupy a substantially different norm regime from the input embeddings the model was trained to"},"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":"2605.12374","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2026-05-12T16:41:09Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"078b5893e734b337f8900d18e5057ee95382897a63ea6c62ef0eb98bba77b3e1","abstract_canon_sha256":"10d2f7dd43887f0a044320abd6aca13826cc64a402c84c55fb30a8ee5b52eaa4"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:00:43.192371Z","signature_b64":"HUYT+KTfXAzT+QZ3EElqzn/huf+JSdDEYGwOfzm9YqLv29gf2+8Lh9BKseNXA8LERT14C6vGxG5LSNP9rinBBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4e81d219658a62ac0c0013f04f00229b27cb8e84219c055bab85f053938dad0f","last_reissued_at":"2026-05-20T00:00:43.191487Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:00:43.191487Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Granular alignment at three levels lets MLLMs generate stable visual latents by fixing decoder-to-input mismatch.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CV","authors_text":"Dexin Wang, Guanjun Jiang, Hao Li, Lei Lv, Li Wang, Mengyu Zhou, Pascal Poupart, Qi Zhao, Xiaoxi Jiang, Yanting Miao, Yutao Sun","submitted_at":"2026-05-12T16:41:09Z","abstract_excerpt":"Visual latent reasoning lets a multimodal large language model (MLLM) create intermediate visual evidence as continuous tokens, avoiding external tools or image generators. However, existing methods usually follow an output-as-input latent paradigm and yield unstable gains. We identify evidence for a feature-space mismatch that can contribute to this instability: dominant visual-latent models build on pre-norm MLLMs and reuse decoder hidden states as predicted latent inputs, even though these states occupy a substantially different norm regime from the input embeddings the model was trained to"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On Qwen2.5-VL 7B, the resulting model achieves the best mean aggregate perception and reasoning performance among our supervised variants. Inference-time intervention probing further suggests that generated latents provide task-relevant visual signal beyond merely adding token slots.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The feature-space mismatch between decoder hidden states and input embeddings in pre-norm MLLMs is a primary contributor to instability in existing output-as-input visual-latent methods.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GAP introduces three-level alignment for visual latent reasoning in MLLMs, achieving top aggregate perception and reasoning performance on Qwen2.5-VL 7B by addressing decoder-input norm mismatch.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Granular alignment at three levels lets MLLMs generate stable visual latents by fixing decoder-to-input mismatch.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"312c089eecf315deb59a4b908c7c1d72903862874c4b1a7e7c45daffb66607c7"},"source":{"id":"2605.12374","kind":"arxiv","version":2},"verdict":{"id":"02fb4ab2-0e36-4c4a-ace2-4ee9540cf8e4","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T16:43:15.588650Z","strongest_claim":"On Qwen2.5-VL 7B, the resulting model achieves the best mean aggregate perception and reasoning performance among our supervised variants. Inference-time intervention probing further suggests that generated latents provide task-relevant visual signal beyond merely adding token slots.","one_line_summary":"GAP introduces three-level alignment for visual latent reasoning in MLLMs, achieving top aggregate perception and reasoning performance on Qwen2.5-VL 7B by addressing decoder-input norm mismatch.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The feature-space mismatch between decoder hidden states and input embeddings in pre-norm MLLMs is a primary contributor to instability in existing output-as-input visual-latent methods.","pith_extraction_headline":"Granular alignment at three levels lets MLLMs generate stable visual latents by fixing decoder-to-input mismatch."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.12374/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-19T22:41:58.233759Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T10:37:07.135846Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T08:01:18.885772Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T07:36:34.939531Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"7ca5a5ba5f26fc82ef388b31edef12f3b35ef689167d37d2524c2adb8416869b"},"references":{"count":30,"sample":[{"doi":"","year":null,"title":"Imagine while Reasoning in Space: Multimodal Visualization-of-Thought","work_id":"fe549c4a-19fa-421a-9a30-230932ba737e","ref_index":1,"cited_arxiv_id":"2501.07542","is_internal_anchor":true},{"doi":"","year":null,"title":"arXiv preprint arXiv:2510.24514 , year=","work_id":"b88a3f18-3bbc-44da-82ee-292b974b146d","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Gemma 3 Technical Report","work_id":"f93e08bf-9e96-409b-8ac6-b8385fd17fd7","ref_index":3,"cited_arxiv_id":"2503.19786","is_internal_anchor":true},{"doi":"","year":null,"title":"OpenAI GPT-5 System Card","work_id":"ca87689a-0d29-4476-b504-b65dbbb08af4","ref_index":4,"cited_arxiv_id":"2601.03267","is_internal_anchor":true},{"doi":"","year":2025,"title":"Qwen3-VL Technical Report , author=. 2025 , eprint=","work_id":"a6efb89e-81e1-4208-bbf7-b8686a399f3b","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":30,"snapshot_sha256":"b9f34aa8da27cf1d3e7850ed1bb1604d2ec924fe046279650c06afb37704faef","internal_anchors":16},"formal_canon":{"evidence_count":2,"snapshot_sha256":"a5f1b36fc9fe7c5839b360e70d1c355e19646d8baa9ffe851ed75936ef692891"},"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":"2605.12374","created_at":"2026-05-20T00:00:43.191637+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.12374v2","created_at":"2026-05-20T00:00:43.191637+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.12374","created_at":"2026-05-20T00:00:43.191637+00:00"},{"alias_kind":"pith_short_12","alias_value":"J2A5EGLFRJRK","created_at":"2026-05-20T00:00:43.191637+00:00"},{"alias_kind":"pith_short_16","alias_value":"J2A5EGLFRJRKYDAA","created_at":"2026-05-20T00:00:43.191637+00:00"},{"alias_kind":"pith_short_8","alias_value":"J2A5EGLF","created_at":"2026-05-20T00:00:43.191637+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/J2A5EGLFRJRKYDAACPYE6ABCTM","json":"https://pith.science/pith/J2A5EGLFRJRKYDAACPYE6ABCTM.json","graph_json":"https://pith.science/api/pith-number/J2A5EGLFRJRKYDAACPYE6ABCTM/graph.json","events_json":"https://pith.science/api/pith-number/J2A5EGLFRJRKYDAACPYE6ABCTM/events.json","paper":"https://pith.science/paper/J2A5EGLF"},"agent_actions":{"view_html":"https://pith.science/pith/J2A5EGLFRJRKYDAACPYE6ABCTM","download_json":"https://pith.science/pith/J2A5EGLFRJRKYDAACPYE6ABCTM.json","view_paper":"https://pith.science/paper/J2A5EGLF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.12374&json=true","fetch_graph":"https://pith.science/api/pith-number/J2A5EGLFRJRKYDAACPYE6ABCTM/graph.json","fetch_events":"https://pith.science/api/pith-number/J2A5EGLFRJRKYDAACPYE6ABCTM/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/J2A5EGLFRJRKYDAACPYE6ABCTM/action/timestamp_anchor","attest_storage":"https://pith.science/pith/J2A5EGLFRJRKYDAACPYE6ABCTM/action/storage_attestation","attest_author":"https://pith.science/pith/J2A5EGLFRJRKYDAACPYE6ABCTM/action/author_attestation","sign_citation":"https://pith.science/pith/J2A5EGLFRJRKYDAACPYE6ABCTM/action/citation_signature","submit_replication":"https://pith.science/pith/J2A5EGLFRJRKYDAACPYE6ABCTM/action/replication_record"}},"created_at":"2026-05-20T00:00:43.191637+00:00","updated_at":"2026-05-20T00:00:43.191637+00:00"}