{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:S7VVMMSIJIWELJMQ6JJ7ULDKS2","short_pith_number":"pith:S7VVMMSI","schema_version":"1.0","canonical_sha256":"97eb5632484a2c45a590f253fa2c6a96836ddcb499a0d45486327a6bf1c14baf","source":{"kind":"arxiv","id":"2406.03496","version":1},"attestation_state":"computed","paper":{"title":"Wings: Learning Multimodal LLMs without Text-only Forgetting","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"De-Chuan Zhan, Han-Jia Ye, Kaifu Zhang, Qing-Guo Chen, Shiyin Lu, Weihua Luo, Yang Li, Yanqing Ma, Yi-Kai Zhang, Zhao Xu","submitted_at":"2024-06-05T17:59:40Z","abstract_excerpt":"Multimodal large language models (MLLMs), initiated with a trained LLM, first align images with text and then fine-tune on multimodal mixed inputs. However, the MLLM catastrophically forgets the text-only instructions, which do not include images and can be addressed within the initial LLM. In this paper, we present Wings, a novel MLLM that excels in both text-only dialogues and multimodal comprehension. Analyzing MLLM attention in multimodal instructions reveals that text-only forgetting is related to the attention shifts from pre-image to post-image text. From that, we construct extra module"},"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":"2406.03496","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-06-05T17:59:40Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"8ff0d71f5f0bb14daae58507aa629c86e7f5370866bdca8bf7a85725e363b1dd","abstract_canon_sha256":"0150563e79b100616c8eb00b2ac02a9864c67d53ec79342ccdc0543cf6860816"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:27:58.473208Z","signature_b64":"Hp2mcW1qiVhwclq4GxR5RUg+CM1w4f4OStcnnN0x+o33+oHw6NZnu3psLDguxdFx+8ZCflEifou7PwReZMiSCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"97eb5632484a2c45a590f253fa2c6a96836ddcb499a0d45486327a6bf1c14baf","last_reissued_at":"2026-07-05T08:27:58.472724Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:27:58.472724Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Wings: Learning Multimodal LLMs without Text-only Forgetting","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"De-Chuan Zhan, Han-Jia Ye, Kaifu Zhang, Qing-Guo Chen, Shiyin Lu, Weihua Luo, Yang Li, Yanqing Ma, Yi-Kai Zhang, Zhao Xu","submitted_at":"2024-06-05T17:59:40Z","abstract_excerpt":"Multimodal large language models (MLLMs), initiated with a trained LLM, first align images with text and then fine-tune on multimodal mixed inputs. However, the MLLM catastrophically forgets the text-only instructions, which do not include images and can be addressed within the initial LLM. In this paper, we present Wings, a novel MLLM that excels in both text-only dialogues and multimodal comprehension. Analyzing MLLM attention in multimodal instructions reveals that text-only forgetting is related to the attention shifts from pre-image to post-image text. From that, we construct extra module"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.03496","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/2406.03496/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":"2406.03496","created_at":"2026-07-05T08:27:58.472772+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.03496v1","created_at":"2026-07-05T08:27:58.472772+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.03496","created_at":"2026-07-05T08:27:58.472772+00:00"},{"alias_kind":"pith_short_12","alias_value":"S7VVMMSIJIWE","created_at":"2026-07-05T08:27:58.472772+00:00"},{"alias_kind":"pith_short_16","alias_value":"S7VVMMSIJIWELJMQ","created_at":"2026-07-05T08:27:58.472772+00:00"},{"alias_kind":"pith_short_8","alias_value":"S7VVMMSI","created_at":"2026-07-05T08:27:58.472772+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2506.16691","citing_title":"LaVi: Efficient Large Vision-Language Models via Internal Feature Modulation","ref_index":72,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/S7VVMMSIJIWELJMQ6JJ7ULDKS2","json":"https://pith.science/pith/S7VVMMSIJIWELJMQ6JJ7ULDKS2.json","graph_json":"https://pith.science/api/pith-number/S7VVMMSIJIWELJMQ6JJ7ULDKS2/graph.json","events_json":"https://pith.science/api/pith-number/S7VVMMSIJIWELJMQ6JJ7ULDKS2/events.json","paper":"https://pith.science/paper/S7VVMMSI"},"agent_actions":{"view_html":"https://pith.science/pith/S7VVMMSIJIWELJMQ6JJ7ULDKS2","download_json":"https://pith.science/pith/S7VVMMSIJIWELJMQ6JJ7ULDKS2.json","view_paper":"https://pith.science/paper/S7VVMMSI","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.03496&json=true","fetch_graph":"https://pith.science/api/pith-number/S7VVMMSIJIWELJMQ6JJ7ULDKS2/graph.json","fetch_events":"https://pith.science/api/pith-number/S7VVMMSIJIWELJMQ6JJ7ULDKS2/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/S7VVMMSIJIWELJMQ6JJ7ULDKS2/action/timestamp_anchor","attest_storage":"https://pith.science/pith/S7VVMMSIJIWELJMQ6JJ7ULDKS2/action/storage_attestation","attest_author":"https://pith.science/pith/S7VVMMSIJIWELJMQ6JJ7ULDKS2/action/author_attestation","sign_citation":"https://pith.science/pith/S7VVMMSIJIWELJMQ6JJ7ULDKS2/action/citation_signature","submit_replication":"https://pith.science/pith/S7VVMMSIJIWELJMQ6JJ7ULDKS2/action/replication_record"}},"created_at":"2026-07-05T08:27:58.472772+00:00","updated_at":"2026-07-05T08:27:58.472772+00:00"}