{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:JVOVEPHLNRHC3WTAD5GEUUDJH2","short_pith_number":"pith:JVOVEPHL","schema_version":"1.0","canonical_sha256":"4d5d523ceb6c4e2dda601f4c4a50693e968c27ca5de481f06d55addca6602847","source":{"kind":"arxiv","id":"2605.13156","version":1},"attestation_state":"computed","paper":{"title":"Dual-Pathway Circuits of Object Hallucination in Vision-Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Vision-language models contain a distinct hallucination pathway that can be suppressed to cut object errors by up to 76 percent with little accuracy loss.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Aofan Liu, Ding Zhong, Guangyuan Dong, Jiaxin Liu, Pengcheng Fang, Qishi Zhan, Yue Wang, Zhaolu Kang, Zhidong Yang","submitted_at":"2026-05-13T08:20:01Z","abstract_excerpt":"Vision-language models (VLMs) have demonstrated remarkable capabilities in bridging visual perception and natural language understanding, enabling a wide range of multimodal reasoning tasks. However, they often produce object hallucinations, describing content absent from the input image, which limits their reliability and interpretability. To address this limitation, we propose Dual-Pathway Circuit Analysis, a framework that identifies and characterizes hallucination-related circuits in VLMs for mechanistic understanding and causal probing. We first apply activation patching across five archi"},"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":false},"canonical_record":{"source":{"id":"2605.13156","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2026-05-13T08:20:01Z","cross_cats_sorted":[],"title_canon_sha256":"f89aa4ccbc871b468ff0db0c36bd2b0a0058586af4a5bf4f5f1227a05baba620","abstract_canon_sha256":"fb6634c051d1f47045c821ce868cb73fcd37c04fcc950a252aee8ec9a350994c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:08:57.001341Z","signature_b64":"jnu3Edg6INvxf9VnQKP3RjElpqP5ZwBGrzYLp47zAR+auV3PeApdCw3HRPuj1pwBmEhIVZk+3xWopZlKsWl5Cg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4d5d523ceb6c4e2dda601f4c4a50693e968c27ca5de481f06d55addca6602847","last_reissued_at":"2026-05-18T03:08:57.000692Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:08:57.000692Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Dual-Pathway Circuits of Object Hallucination in Vision-Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Vision-language models contain a distinct hallucination pathway that can be suppressed to cut object errors by up to 76 percent with little accuracy loss.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Aofan Liu, Ding Zhong, Guangyuan Dong, Jiaxin Liu, Pengcheng Fang, Qishi Zhan, Yue Wang, Zhaolu Kang, Zhidong Yang","submitted_at":"2026-05-13T08:20:01Z","abstract_excerpt":"Vision-language models (VLMs) have demonstrated remarkable capabilities in bridging visual perception and natural language understanding, enabling a wide range of multimodal reasoning tasks. However, they often produce object hallucinations, describing content absent from the input image, which limits their reliability and interpretability. To address this limitation, we propose Dual-Pathway Circuit Analysis, a framework that identifies and characterizes hallucination-related circuits in VLMs for mechanistic understanding and causal probing. We first apply activation patching across five archi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"targeted suppression of hallucination-pathway components, showing that scaling these components reduces object hallucination by up to 76% with minimal accuracy cost, and validate that the same circuit selectively transfers to relational but not attribute hallucination","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That activation patching and the observed polarity flip in grounding components causally identify and control hallucination behavior rather than reflecting correlated but non-causal patterns in model activations.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Vision-language models contain identifiable grounding and hallucination pathways; suppressing the latter reduces object hallucinations by up to 76% while preserving accuracy.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Vision-language models contain a distinct hallucination pathway that can be suppressed to cut object errors by up to 76 percent with little accuracy loss.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"cb73c54f77890730c0e1932b57fdc0580102953516efa8c18d7e9efe1d911645"},"source":{"id":"2605.13156","kind":"arxiv","version":1},"verdict":{"id":"437517d2-f5c8-493c-8f61-f80c30584198","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:14:39.409389Z","strongest_claim":"targeted suppression of hallucination-pathway components, showing that scaling these components reduces object hallucination by up to 76% with minimal accuracy cost, and validate that the same circuit selectively transfers to relational but not attribute hallucination","one_line_summary":"Vision-language models contain identifiable grounding and hallucination pathways; suppressing the latter reduces object hallucinations by up to 76% while preserving accuracy.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That activation patching and the observed polarity flip in grounding components causally identify and control hallucination behavior rather than reflecting correlated but non-causal patterns in model activations.","pith_extraction_headline":"Vision-language models contain a distinct hallucination pathway that can be suppressed to cut object errors by up to 76 percent with little accuracy loss."},"references":{"count":42,"sample":[{"doi":"","year":2025,"title":"Qwen3-VL Technical Report","work_id":"1fe243aa-e3c0-4da6-b391-4cbcfc88d5c0","ref_index":1,"cited_arxiv_id":"2511.21631","is_internal_anchor":true},{"doi":"","year":2025,"title":"Qwen2.5-VL Technical Report","work_id":"69dffacb-bfe8-442d-be86-48624c60426f","ref_index":2,"cited_arxiv_id":"2502.13923","is_internal_anchor":true},{"doi":"","year":2025,"title":"Univg-r1: Reasoning guided universal visual grounding with reinforcement learning","work_id":"fad4943c-f29d-4daa-a7ba-dd6f03aac5aa","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"A survey of multimodal hallucination evaluation and detection.International Journal of Computer Vision, 2025","work_id":"1f934e4b-f329-42cc-a67a-e40c379c2501","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Towards automated circuit discovery for mechanistic interpretability.Advances in Neural Information Processing Systems, 36, 2023","work_id":"3e26f32d-04ea-470a-82eb-37cb38d6c113","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":42,"snapshot_sha256":"28840d77e71dcf3c5663726ecca153a82326e8ba51aeaf073cf5c06b136ff7ae","internal_anchors":14},"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":"2605.13156","created_at":"2026-05-18T03:08:57.000790+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.13156v1","created_at":"2026-05-18T03:08:57.000790+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13156","created_at":"2026-05-18T03:08:57.000790+00:00"},{"alias_kind":"pith_short_12","alias_value":"JVOVEPHLNRHC","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"JVOVEPHLNRHC3WTA","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"JVOVEPHL","created_at":"2026-05-18T12:33:37.589309+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/JVOVEPHLNRHC3WTAD5GEUUDJH2","json":"https://pith.science/pith/JVOVEPHLNRHC3WTAD5GEUUDJH2.json","graph_json":"https://pith.science/api/pith-number/JVOVEPHLNRHC3WTAD5GEUUDJH2/graph.json","events_json":"https://pith.science/api/pith-number/JVOVEPHLNRHC3WTAD5GEUUDJH2/events.json","paper":"https://pith.science/paper/JVOVEPHL"},"agent_actions":{"view_html":"https://pith.science/pith/JVOVEPHLNRHC3WTAD5GEUUDJH2","download_json":"https://pith.science/pith/JVOVEPHLNRHC3WTAD5GEUUDJH2.json","view_paper":"https://pith.science/paper/JVOVEPHL","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.13156&json=true","fetch_graph":"https://pith.science/api/pith-number/JVOVEPHLNRHC3WTAD5GEUUDJH2/graph.json","fetch_events":"https://pith.science/api/pith-number/JVOVEPHLNRHC3WTAD5GEUUDJH2/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/JVOVEPHLNRHC3WTAD5GEUUDJH2/action/timestamp_anchor","attest_storage":"https://pith.science/pith/JVOVEPHLNRHC3WTAD5GEUUDJH2/action/storage_attestation","attest_author":"https://pith.science/pith/JVOVEPHLNRHC3WTAD5GEUUDJH2/action/author_attestation","sign_citation":"https://pith.science/pith/JVOVEPHLNRHC3WTAD5GEUUDJH2/action/citation_signature","submit_replication":"https://pith.science/pith/JVOVEPHLNRHC3WTAD5GEUUDJH2/action/replication_record"}},"created_at":"2026-05-18T03:08:57.000790+00:00","updated_at":"2026-05-18T03:08:57.000790+00:00"}