{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:ZWGVMJJ2JX25VTRNLHTHWSOTAQ","short_pith_number":"pith:ZWGVMJJ2","schema_version":"1.0","canonical_sha256":"cd8d56253a4df5dace2d59e67b49d30413b25394b772eef5b3af02d07fb30d2c","source":{"kind":"arxiv","id":"2504.00557","version":1},"attestation_state":"computed","paper":{"title":"Efficient LLaMA-3.2-Vision by Trimming Cross-attended Visual Features","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CV","authors_text":"Bo-Kyeong Kim, Donguk Lim, Jaeyeon Kim, Jewon Lee, Ki-Ung Song, Seungmin Yang, Tae-Ho Kim, Wooksu Shin, Yong Jae Lee","submitted_at":"2025-04-01T09:10:32Z","abstract_excerpt":"Visual token reduction lowers inference costs caused by extensive image features in large vision-language models (LVLMs). Unlike relevant studies that prune tokens in self-attention-only LVLMs, our work uniquely addresses cross-attention-based models, which achieve superior performance. We identify that the key-value (KV) cache size for image tokens in cross-attention layers significantly exceeds that of text tokens in self-attention layers, posing a major compute bottleneck. To mitigate this issue, we exploit the sparse nature in cross-attention maps to selectively prune redundant visual feat"},"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":"2504.00557","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2025-04-01T09:10:32Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"2a7b2c9261d1749c6db593329eba7e995b781757d8963f83385d608d7189cb86","abstract_canon_sha256":"226ca81ba93b120c0d1259a9cff41f8c94deb70705d8cfbd6f2817bff843b10e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:42:47.016555Z","signature_b64":"xCOpue2fd0wnTy/bLpSUdQM3icXeUEDN56vTPxvaEUkuTU0Kdwg/j5ISFkgJHW9xDf8d4EMM94DcCBkBBUInDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"cd8d56253a4df5dace2d59e67b49d30413b25394b772eef5b3af02d07fb30d2c","last_reissued_at":"2026-07-05T10:42:47.016122Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:42:47.016122Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Efficient LLaMA-3.2-Vision by Trimming Cross-attended Visual Features","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CV","authors_text":"Bo-Kyeong Kim, Donguk Lim, Jaeyeon Kim, Jewon Lee, Ki-Ung Song, Seungmin Yang, Tae-Ho Kim, Wooksu Shin, Yong Jae Lee","submitted_at":"2025-04-01T09:10:32Z","abstract_excerpt":"Visual token reduction lowers inference costs caused by extensive image features in large vision-language models (LVLMs). Unlike relevant studies that prune tokens in self-attention-only LVLMs, our work uniquely addresses cross-attention-based models, which achieve superior performance. We identify that the key-value (KV) cache size for image tokens in cross-attention layers significantly exceeds that of text tokens in self-attention layers, posing a major compute bottleneck. To mitigate this issue, we exploit the sparse nature in cross-attention maps to selectively prune redundant visual feat"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.00557","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/2504.00557/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":"2504.00557","created_at":"2026-07-05T10:42:47.016186+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.00557v1","created_at":"2026-07-05T10:42:47.016186+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.00557","created_at":"2026-07-05T10:42:47.016186+00:00"},{"alias_kind":"pith_short_12","alias_value":"ZWGVMJJ2JX25","created_at":"2026-07-05T10:42:47.016186+00:00"},{"alias_kind":"pith_short_16","alias_value":"ZWGVMJJ2JX25VTRN","created_at":"2026-07-05T10:42:47.016186+00:00"},{"alias_kind":"pith_short_8","alias_value":"ZWGVMJJ2","created_at":"2026-07-05T10:42:47.016186+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.06626","citing_title":"Reward Valuation in Vision Language Models: Causal Mechanisms Underlying Anhedonia","ref_index":114,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10744","citing_title":"C-CoT: Counterfactual Chain-of-Thought with Vision-Language Models for Safe Autonomous Driving","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2604.25614","citing_title":"HotComment: A Benchmark for Evaluating Popularity of Online Comments","ref_index":36,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19083","citing_title":"ProjLens: Unveiling the Role of Projectors in Multimodal Model Safety","ref_index":44,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ZWGVMJJ2JX25VTRNLHTHWSOTAQ","json":"https://pith.science/pith/ZWGVMJJ2JX25VTRNLHTHWSOTAQ.json","graph_json":"https://pith.science/api/pith-number/ZWGVMJJ2JX25VTRNLHTHWSOTAQ/graph.json","events_json":"https://pith.science/api/pith-number/ZWGVMJJ2JX25VTRNLHTHWSOTAQ/events.json","paper":"https://pith.science/paper/ZWGVMJJ2"},"agent_actions":{"view_html":"https://pith.science/pith/ZWGVMJJ2JX25VTRNLHTHWSOTAQ","download_json":"https://pith.science/pith/ZWGVMJJ2JX25VTRNLHTHWSOTAQ.json","view_paper":"https://pith.science/paper/ZWGVMJJ2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.00557&json=true","fetch_graph":"https://pith.science/api/pith-number/ZWGVMJJ2JX25VTRNLHTHWSOTAQ/graph.json","fetch_events":"https://pith.science/api/pith-number/ZWGVMJJ2JX25VTRNLHTHWSOTAQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ZWGVMJJ2JX25VTRNLHTHWSOTAQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ZWGVMJJ2JX25VTRNLHTHWSOTAQ/action/storage_attestation","attest_author":"https://pith.science/pith/ZWGVMJJ2JX25VTRNLHTHWSOTAQ/action/author_attestation","sign_citation":"https://pith.science/pith/ZWGVMJJ2JX25VTRNLHTHWSOTAQ/action/citation_signature","submit_replication":"https://pith.science/pith/ZWGVMJJ2JX25VTRNLHTHWSOTAQ/action/replication_record"}},"created_at":"2026-07-05T10:42:47.016186+00:00","updated_at":"2026-07-05T10:42:47.016186+00:00"}