{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:GEJ4EWA452CQZNMFHTM2FDOQXE","short_pith_number":"pith:GEJ4EWA4","schema_version":"1.0","canonical_sha256":"3113c2581cee850cb5853cd9a28dd0b93b7f21fc377a52ef6fcc151688c43c5c","source":{"kind":"arxiv","id":"2506.14766","version":3},"attestation_state":"computed","paper":{"title":"ASCD: Attention-Steerable Contrastive Decoding for Reducing Hallucination in MLLM","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.CV","authors_text":"Aniri, Jinhe Bi, Soeren Pirk, Yujun Wang, Yunpu Ma","submitted_at":"2025-06-17T17:58:11Z","abstract_excerpt":"Multimodal large language models (MLLMs) frequently hallucinate by over-committing to spurious visual cues. Prior remedies-Visual and Instruction Contrastive Decoding (VCD, ICD)-mitigate this issue, yet the mechanism remains opaque. We first empirically show that their improvements systematically coincide with redistributions of cross-modal attention. Building on this insight, we propose Attention-Steerable Contrastive Decoding (ASCD), which directly steers the attention scores during decoding. ASCD combines (i) positive steering, which amplifies automatically mined text-centric heads-stable w"},"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":"2506.14766","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2025-06-17T17:58:11Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"92ac4ac864f0d4ba333f2bacb00084286999146aebe4e079b99dc4e0d76744c1","abstract_canon_sha256":"b5ef09606da4d40d03cf18c8e4309250e2097b9f036a5214891dc3b10f18b90f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3113c2581cee850cb5853cd9a28dd0b93b7f21fc377a52ef6fcc151688c43c5c","last_reissued_at":"2026-07-30T01:18:57.550140Z","signature_status":"unsigned_v0","first_computed_at":"2026-07-30T01:18:57.550140Z"},"graph_snapshot":{"paper":{"title":"ASCD: Attention-Steerable Contrastive Decoding for Reducing Hallucination in MLLM","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.CV","authors_text":"Aniri, Jinhe Bi, Soeren Pirk, Yujun Wang, Yunpu Ma","submitted_at":"2025-06-17T17:58:11Z","abstract_excerpt":"Multimodal large language models (MLLMs) frequently hallucinate by over-committing to spurious visual cues. Prior remedies-Visual and Instruction Contrastive Decoding (VCD, ICD)-mitigate this issue, yet the mechanism remains opaque. We first empirically show that their improvements systematically coincide with redistributions of cross-modal attention. Building on this insight, we propose Attention-Steerable Contrastive Decoding (ASCD), which directly steers the attention scores during decoding. ASCD combines (i) positive steering, which amplifies automatically mined text-centric heads-stable w"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.14766","kind":"arxiv","version":3},"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/2506.14766/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":"2506.14766","created_at":"2026-07-30T01:18:57.555331+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.14766v3","created_at":"2026-07-30T01:18:57.555331+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.14766","created_at":"2026-07-30T01:18:57.555331+00:00"},{"alias_kind":"pith_short_12","alias_value":"GEJ4EWA452CQ","created_at":"2026-07-30T01:18:57.555331+00:00"},{"alias_kind":"pith_short_16","alias_value":"GEJ4EWA452CQZNMF","created_at":"2026-07-30T01:18:57.555331+00:00"},{"alias_kind":"pith_short_8","alias_value":"GEJ4EWA4","created_at":"2026-07-30T01:18:57.555331+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":12,"internal_anchor_count":12,"sample":[{"citing_arxiv_id":"2606.20419","citing_title":"Spectral Query-Key Product Weight Steering for Training-Free VLM Hallucination Mitigation","ref_index":37,"is_internal_anchor":true},{"citing_arxiv_id":"2606.17953","citing_title":"MLLMs Get It Right, Then Get It Wrong: Tracing and Correcting Late-Layer Textual Bias","ref_index":33,"is_internal_anchor":true},{"citing_arxiv_id":"2605.27993","citing_title":"Rethinking Visual Neglect: Steering via Context-Preference for MLLM Hallucination Mitigation","ref_index":26,"is_internal_anchor":true},{"citing_arxiv_id":"2510.19316","citing_title":"KORE: Enhancing Knowledge Injection for Large Multimodal Models via Knowledge-Oriented Controls","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2604.26614","citing_title":"State Beyond Appearance: Diagnosing and Improving State Consistency in Dial-Based Measurement Reading","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2604.19386","citing_title":"Air-Know: Arbiter-Calibrated Knowledge-Internalizing Robust Network for Composed Image Retrieval","ref_index":43,"is_internal_anchor":true},{"citing_arxiv_id":"2604.09749","citing_title":"See Fair, Speak Truth: Equitable Attention Improves Grounding and Reduces Hallucination in Vision-Language Alignment","ref_index":26,"is_internal_anchor":true},{"citing_arxiv_id":"2605.07817","citing_title":"GazeVLM: Active Vision via Internal Attention Control for Multimodal Reasoning","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2604.17898","citing_title":"ReTrack: Evidence-Driven Dual-Stream Directional Anchor Calibration Network for Composed Video Retrieval","ref_index":67,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18051","citing_title":"INTENT: Invariance and Discrimination-aware Noise Mitigation for Robust Composed Image Retrieval","ref_index":89,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18037","citing_title":"HABIT: Chrono-Synergia Robust Progressive Learning Framework for Composed Image Retrieval","ref_index":95,"is_internal_anchor":true},{"citing_arxiv_id":"2604.20358","citing_title":"ConeSep: Cone-based Robust Noise-Unlearning Compositional Network for Composed Image Retrieval","ref_index":54,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/GEJ4EWA452CQZNMFHTM2FDOQXE","json":"https://pith.science/pith/GEJ4EWA452CQZNMFHTM2FDOQXE.json","graph_json":"https://pith.science/api/pith-number/GEJ4EWA452CQZNMFHTM2FDOQXE/graph.json","events_json":"https://pith.science/api/pith-number/GEJ4EWA452CQZNMFHTM2FDOQXE/events.json","paper":"https://pith.science/paper/GEJ4EWA4"},"agent_actions":{"view_html":"https://pith.science/pith/GEJ4EWA452CQZNMFHTM2FDOQXE","download_json":"https://pith.science/pith/GEJ4EWA452CQZNMFHTM2FDOQXE.json","view_paper":"https://pith.science/paper/GEJ4EWA4","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.14766&json=true","fetch_graph":"https://pith.science/api/pith-number/GEJ4EWA452CQZNMFHTM2FDOQXE/graph.json","fetch_events":"https://pith.science/api/pith-number/GEJ4EWA452CQZNMFHTM2FDOQXE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/GEJ4EWA452CQZNMFHTM2FDOQXE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/GEJ4EWA452CQZNMFHTM2FDOQXE/action/storage_attestation","attest_author":"https://pith.science/pith/GEJ4EWA452CQZNMFHTM2FDOQXE/action/author_attestation","sign_citation":"https://pith.science/pith/GEJ4EWA452CQZNMFHTM2FDOQXE/action/citation_signature","submit_replication":"https://pith.science/pith/GEJ4EWA452CQZNMFHTM2FDOQXE/action/replication_record"}},"created_at":"2026-07-30T01:18:57.555331+00:00","updated_at":"2026-07-30T01:18:57.555331+00:00"}