{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:563GZFCFEZKCIJURH2NJHMT5VB","short_pith_number":"pith:563GZFCF","schema_version":"1.0","canonical_sha256":"efb66c944526542426913e9a93b27da87a3698fcdc78b91d37d59c62dbcdfb6b","source":{"kind":"arxiv","id":"2407.02333","version":1},"attestation_state":"computed","paper":{"title":"Why do LLaVA Vision-Language Models Reply to Images in English?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CV"],"primary_cat":"cs.CL","authors_text":"Anahita Bhiwandiwalla, Anne Lauscher, Carolin Holtermann, Florian Schneider, Matthew Lyle Olson, Musashi Hinck, Shaoyen Tseng, Sungduk Yu, Vasudev Lal","submitted_at":"2024-07-02T15:01:55Z","abstract_excerpt":"We uncover a surprising multilingual bias occurring in a popular class of multimodal vision-language models (VLMs). Including an image in the query to a LLaVA-style VLM significantly increases the likelihood of the model returning an English response, regardless of the language of the query. This paper investigates the causes of this loss with a two-pronged approach that combines extensive ablation of the design space with a mechanistic analysis of the models' internal representations of image and text inputs. Both approaches indicate that the issue stems in the language modelling component of"},"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":"2407.02333","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-07-02T15:01:55Z","cross_cats_sorted":["cs.CV"],"title_canon_sha256":"a377eff375c1d10ef66388c47033dc7b9cd697db0aff551e9a1c4a83cbc1c44d","abstract_canon_sha256":"e0bd3abf0cac252012b1e5dedbc45c81ef6000de8b84fd82cd03d4cfa3929448"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:39:22.830598Z","signature_b64":"0+GCHEsNQyrRMYxua4syYrpFho4CrDJh5UrIBKCwbzbSIpLYOvUdvZ5fdDIVOYnwhjef9NMNLV8taAtRBSwqCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"efb66c944526542426913e9a93b27da87a3698fcdc78b91d37d59c62dbcdfb6b","last_reissued_at":"2026-07-05T08:39:22.830094Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:39:22.830094Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Why do LLaVA Vision-Language Models Reply to Images in English?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CV"],"primary_cat":"cs.CL","authors_text":"Anahita Bhiwandiwalla, Anne Lauscher, Carolin Holtermann, Florian Schneider, Matthew Lyle Olson, Musashi Hinck, Shaoyen Tseng, Sungduk Yu, Vasudev Lal","submitted_at":"2024-07-02T15:01:55Z","abstract_excerpt":"We uncover a surprising multilingual bias occurring in a popular class of multimodal vision-language models (VLMs). Including an image in the query to a LLaVA-style VLM significantly increases the likelihood of the model returning an English response, regardless of the language of the query. This paper investigates the causes of this loss with a two-pronged approach that combines extensive ablation of the design space with a mechanistic analysis of the models' internal representations of image and text inputs. Both approaches indicate that the issue stems in the language modelling component of"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2407.02333","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/2407.02333/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":"2407.02333","created_at":"2026-07-05T08:39:22.830159+00:00"},{"alias_kind":"arxiv_version","alias_value":"2407.02333v1","created_at":"2026-07-05T08:39:22.830159+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2407.02333","created_at":"2026-07-05T08:39:22.830159+00:00"},{"alias_kind":"pith_short_12","alias_value":"563GZFCFEZKC","created_at":"2026-07-05T08:39:22.830159+00:00"},{"alias_kind":"pith_short_16","alias_value":"563GZFCFEZKCIJUR","created_at":"2026-07-05T08:39:22.830159+00:00"},{"alias_kind":"pith_short_8","alias_value":"563GZFCF","created_at":"2026-07-05T08:39:22.830159+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/563GZFCFEZKCIJURH2NJHMT5VB","json":"https://pith.science/pith/563GZFCFEZKCIJURH2NJHMT5VB.json","graph_json":"https://pith.science/api/pith-number/563GZFCFEZKCIJURH2NJHMT5VB/graph.json","events_json":"https://pith.science/api/pith-number/563GZFCFEZKCIJURH2NJHMT5VB/events.json","paper":"https://pith.science/paper/563GZFCF"},"agent_actions":{"view_html":"https://pith.science/pith/563GZFCFEZKCIJURH2NJHMT5VB","download_json":"https://pith.science/pith/563GZFCFEZKCIJURH2NJHMT5VB.json","view_paper":"https://pith.science/paper/563GZFCF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2407.02333&json=true","fetch_graph":"https://pith.science/api/pith-number/563GZFCFEZKCIJURH2NJHMT5VB/graph.json","fetch_events":"https://pith.science/api/pith-number/563GZFCFEZKCIJURH2NJHMT5VB/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/563GZFCFEZKCIJURH2NJHMT5VB/action/timestamp_anchor","attest_storage":"https://pith.science/pith/563GZFCFEZKCIJURH2NJHMT5VB/action/storage_attestation","attest_author":"https://pith.science/pith/563GZFCFEZKCIJURH2NJHMT5VB/action/author_attestation","sign_citation":"https://pith.science/pith/563GZFCFEZKCIJURH2NJHMT5VB/action/citation_signature","submit_replication":"https://pith.science/pith/563GZFCFEZKCIJURH2NJHMT5VB/action/replication_record"}},"created_at":"2026-07-05T08:39:22.830159+00:00","updated_at":"2026-07-05T08:39:22.830159+00:00"}