{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:7N6CREIPRB4LYSXLTPCGUEGIUZ","short_pith_number":"pith:7N6CREIP","canonical_record":{"source":{"id":"2604.20937","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-04-22T13:28:53Z","cross_cats_sorted":[],"title_canon_sha256":"65e3dccf74eeafb890e29347ea5b357dfb0f0432f15233cef394f1d773256bbc","abstract_canon_sha256":"65d424d0d48dac48eb20781d3f6b5d6f28aff77d3d07398902ea93c7b2bc9e62"},"schema_version":"1.0"},"canonical_sha256":"fb7c28910f8878bc4aeb9bc46a10c8a662d12d712bef84422a34f7744acb9e03","source":{"kind":"arxiv","id":"2604.20937","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2604.20937","created_at":"2026-06-23T01:12:07Z"},{"alias_kind":"arxiv_version","alias_value":"2604.20937v2","created_at":"2026-06-23T01:12:07Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.20937","created_at":"2026-06-23T01:12:07Z"},{"alias_kind":"pith_short_12","alias_value":"7N6CREIPRB4L","created_at":"2026-06-23T01:12:07Z"},{"alias_kind":"pith_short_16","alias_value":"7N6CREIPRB4LYSXL","created_at":"2026-06-23T01:12:07Z"},{"alias_kind":"pith_short_8","alias_value":"7N6CREIP","created_at":"2026-06-23T01:12:07Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:7N6CREIPRB4LYSXLTPCGUEGIUZ","target":"record","payload":{"canonical_record":{"source":{"id":"2604.20937","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-04-22T13:28:53Z","cross_cats_sorted":[],"title_canon_sha256":"65e3dccf74eeafb890e29347ea5b357dfb0f0432f15233cef394f1d773256bbc","abstract_canon_sha256":"65d424d0d48dac48eb20781d3f6b5d6f28aff77d3d07398902ea93c7b2bc9e62"},"schema_version":"1.0"},"canonical_sha256":"fb7c28910f8878bc4aeb9bc46a10c8a662d12d712bef84422a34f7744acb9e03","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-23T01:12:07.447411Z","signature_b64":"rJ69bsuBuDNkioyPhz7TFEcY0jaAUD+7HdV+rwQvJxHLcWX2RQQPM2P364isa/nmDdm1/Em+1TRYBbxjwZAGAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"fb7c28910f8878bc4aeb9bc46a10c8a662d12d712bef84422a34f7744acb9e03","last_reissued_at":"2026-06-23T01:12:07.446890Z","signature_status":"signed_v1","first_computed_at":"2026-06-23T01:12:07.446890Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2604.20937","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-23T01:12:07Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"hSw0/seqIUm04WUyZL2F856HmyzBnhC0HNSMDV9tDDEVs/0yedClGqxzMWQqOCmaNI5kO35iWp3hmk0e2GrOBA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-06T02:36:36.368712Z"},"content_sha256":"0e235d5b8810c29b536c25e4d0e9d9021cd236d718967f6b573b19de40538f65","schema_version":"1.0","event_id":"sha256:0e235d5b8810c29b536c25e4d0e9d9021cd236d718967f6b573b19de40538f65"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:7N6CREIPRB4LYSXLTPCGUEGIUZ","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Sink-Token-Aware Pruning for Fine-Grained Video Understanding in Efficient Video LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Sink tokens that attract too much attention cause standard pruning to fail on precise video tasks, but suppressing them with a new score restores performance even at 90% token reduction.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Chanyoung Park, Jinyoung Moon, Jiwan Kim, Julian McAuley, Kibum Kim, Kyle Min, Yueqi Wang","submitted_at":"2026-04-22T13:28:53Z","abstract_excerpt":"Video Large Language Models (Video LLMs) incur high inference latency due to a large number of visual tokens provided to LLMs. To address this, training-free visual token pruning has emerged as a solution to reduce computational costs; however, existing methods are primarily validated on Multiple-Choice Question Answering (MCQA) benchmarks, where coarse-grained cues often suffice. In this work, we reveal that these methods suffer a sharp performance collapse on fine-grained understanding tasks requiring precise visual grounding, such as hallucination evaluation. To explore this gap, we conduct"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our results demonstrate that SToP significantly boosts performance, even when pruning up to 90% of visual tokens.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That sink tokens are the primary driver of the observed performance collapse on fine-grained tasks and that the introduced sink score reliably identifies and suppresses them without discarding critical visual information or introducing new biases.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Sink-Token-aware Pruning (SToP) suppresses semantically uninformative sink tokens during visual token pruning in Video LLMs, boosting fine-grained performance even at 90% pruning rates across hallucination, reasoning, and MCQA benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Sink tokens that attract too much attention cause standard pruning to fail on precise video tasks, but suppressing them with a new score restores performance even at 90% token reduction.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"def5dff3bd388fd6393262ac3bba3b942318dbfb1d426b7921687f22783c0d8b"},"source":{"id":"2604.20937","kind":"arxiv","version":2},"verdict":{"id":"d840a644-2370-4942-a597-639f4c351ab8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T00:40:28.447792Z","strongest_claim":"Our results demonstrate that SToP significantly boosts performance, even when pruning up to 90% of visual tokens.","one_line_summary":"Sink-Token-aware Pruning (SToP) suppresses semantically uninformative sink tokens during visual token pruning in Video LLMs, boosting fine-grained performance even at 90% pruning rates across hallucination, reasoning, and MCQA benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That sink tokens are the primary driver of the observed performance collapse on fine-grained tasks and that the introduced sink score reliably identifies and suppresses them without discarding critical visual information or introducing new biases.","pith_extraction_headline":"Sink tokens that attract too much attention cause standard pruning to fail on precise video tasks, but suppressing them with a new score restores performance even at 90% token reduction."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.20937/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-21T14:37:52.360566Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-20T01:49:51.421970Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"4b4ffdb04aac40d2fa759e00c30b1d1313b921c9d34f9f177e4460577c19e5e1"},"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"},"verdict_id":"d840a644-2370-4942-a597-639f4c351ab8"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-23T01:12:07Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"wLkTTzDWykJgOPDyJje+c9MYfDifwLp3XsrI5XjznrERtwoN/U0wH3J2sJ2WvvqkJEbQj5lds7dsVGZDBXx5Ag==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-06T02:36:36.369349Z"},"content_sha256":"2e9ffd4f8771dfcb1b138a705a69b0f7942bfd5107c14c6dd564583472c0d33a","schema_version":"1.0","event_id":"sha256:2e9ffd4f8771dfcb1b138a705a69b0f7942bfd5107c14c6dd564583472c0d33a"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/7N6CREIPRB4LYSXLTPCGUEGIUZ/bundle.json","state_url":"https://pith.science/pith/7N6CREIPRB4LYSXLTPCGUEGIUZ/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/7N6CREIPRB4LYSXLTPCGUEGIUZ/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-08-06T02:36:36Z","links":{"resolver":"https://pith.science/pith/7N6CREIPRB4LYSXLTPCGUEGIUZ","bundle":"https://pith.science/pith/7N6CREIPRB4LYSXLTPCGUEGIUZ/bundle.json","state":"https://pith.science/pith/7N6CREIPRB4LYSXLTPCGUEGIUZ/state.json","well_known_bundle":"https://pith.science/.well-known/pith/7N6CREIPRB4LYSXLTPCGUEGIUZ/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:7N6CREIPRB4LYSXLTPCGUEGIUZ","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"65d424d0d48dac48eb20781d3f6b5d6f28aff77d3d07398902ea93c7b2bc9e62","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-04-22T13:28:53Z","title_canon_sha256":"65e3dccf74eeafb890e29347ea5b357dfb0f0432f15233cef394f1d773256bbc"},"schema_version":"1.0","source":{"id":"2604.20937","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2604.20937","created_at":"2026-06-23T01:12:07Z"},{"alias_kind":"arxiv_version","alias_value":"2604.20937v2","created_at":"2026-06-23T01:12:07Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.20937","created_at":"2026-06-23T01:12:07Z"},{"alias_kind":"pith_short_12","alias_value":"7N6CREIPRB4L","created_at":"2026-06-23T01:12:07Z"},{"alias_kind":"pith_short_16","alias_value":"7N6CREIPRB4LYSXL","created_at":"2026-06-23T01:12:07Z"},{"alias_kind":"pith_short_8","alias_value":"7N6CREIP","created_at":"2026-06-23T01:12:07Z"}],"graph_snapshots":[{"event_id":"sha256:2e9ffd4f8771dfcb1b138a705a69b0f7942bfd5107c14c6dd564583472c0d33a","target":"graph","created_at":"2026-06-23T01:12:07Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"Our results demonstrate that SToP significantly boosts performance, even when pruning up to 90% of visual tokens."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That sink tokens are the primary driver of the observed performance collapse on fine-grained tasks and that the introduced sink score reliably identifies and suppresses them without discarding critical visual information or introducing new biases."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Sink-Token-aware Pruning (SToP) suppresses semantically uninformative sink tokens during visual token pruning in Video LLMs, boosting fine-grained performance even at 90% pruning rates across hallucination, reasoning, and MCQA benchmarks."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Sink tokens that attract too much attention cause standard pruning to fail on precise video tasks, but suppressing them with a new score restores performance even at 90% token reduction."}],"snapshot_sha256":"def5dff3bd388fd6393262ac3bba3b942318dbfb1d426b7921687f22783c0d8b"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[{"findings_count":0,"name":"ai_meta_artifact","ran_at":"2026-05-21T14:37:52.360566Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"doi_compliance","ran_at":"2026-05-20T01:49:51.421970Z","status":"completed","version":"1.0.0"}],"endpoint":"/pith/2604.20937/integrity.json","findings":[],"snapshot_sha256":"4b4ffdb04aac40d2fa759e00c30b1d1313b921c9d34f9f177e4460577c19e5e1","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Video Large Language Models (Video LLMs) incur high inference latency due to a large number of visual tokens provided to LLMs. To address this, training-free visual token pruning has emerged as a solution to reduce computational costs; however, existing methods are primarily validated on Multiple-Choice Question Answering (MCQA) benchmarks, where coarse-grained cues often suffice. In this work, we reveal that these methods suffer a sharp performance collapse on fine-grained understanding tasks requiring precise visual grounding, such as hallucination evaluation. To explore this gap, we conduct","authors_text":"Chanyoung Park, Jinyoung Moon, Jiwan Kim, Julian McAuley, Kibum Kim, Kyle Min, Yueqi Wang","cross_cats":[],"headline":"Sink tokens that attract too much attention cause standard pruning to fail on precise video tasks, but suppressing them with a new score restores performance even at 90% token reduction.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-04-22T13:28:53Z","title":"Sink-Token-Aware Pruning for Fine-Grained Video Understanding in Efficient Video LLMs"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2604.20937","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-10T00:40:28.447792Z","id":"d840a644-2370-4942-a597-639f4c351ab8","model_set":{"reader":"grok-4.3"},"one_line_summary":"Sink-Token-aware Pruning (SToP) suppresses semantically uninformative sink tokens during visual token pruning in Video LLMs, boosting fine-grained performance even at 90% pruning rates across hallucination, reasoning, and MCQA benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Sink tokens that attract too much attention cause standard pruning to fail on precise video tasks, but suppressing them with a new score restores performance even at 90% token reduction.","strongest_claim":"Our results demonstrate that SToP significantly boosts performance, even when pruning up to 90% of visual tokens.","weakest_assumption":"That sink tokens are the primary driver of the observed performance collapse on fine-grained tasks and that the introduced sink score reliably identifies and suppresses them without discarding critical visual information or introducing new biases."}},"verdict_id":"d840a644-2370-4942-a597-639f4c351ab8"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:0e235d5b8810c29b536c25e4d0e9d9021cd236d718967f6b573b19de40538f65","target":"record","created_at":"2026-06-23T01:12:07Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"65d424d0d48dac48eb20781d3f6b5d6f28aff77d3d07398902ea93c7b2bc9e62","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-04-22T13:28:53Z","title_canon_sha256":"65e3dccf74eeafb890e29347ea5b357dfb0f0432f15233cef394f1d773256bbc"},"schema_version":"1.0","source":{"id":"2604.20937","kind":"arxiv","version":2}},"canonical_sha256":"fb7c28910f8878bc4aeb9bc46a10c8a662d12d712bef84422a34f7744acb9e03","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"fb7c28910f8878bc4aeb9bc46a10c8a662d12d712bef84422a34f7744acb9e03","first_computed_at":"2026-06-23T01:12:07.446890Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-06-23T01:12:07.446890Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"rJ69bsuBuDNkioyPhz7TFEcY0jaAUD+7HdV+rwQvJxHLcWX2RQQPM2P364isa/nmDdm1/Em+1TRYBbxjwZAGAg==","signature_status":"signed_v1","signed_at":"2026-06-23T01:12:07.447411Z","signed_message":"canonical_sha256_bytes"},"source_id":"2604.20937","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:0e235d5b8810c29b536c25e4d0e9d9021cd236d718967f6b573b19de40538f65","sha256:2e9ffd4f8771dfcb1b138a705a69b0f7942bfd5107c14c6dd564583472c0d33a"],"state_sha256":"1c300ea08a6523320add8ee21a4ca92d04e51648b875c9fabdc9fdff9ae681e0"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"UIa3GiAWjXSKi4VfdmQsxZsuuwUp1NMvWX8JtTLvX2fwEna286GTpdCdfv7AIrFwJqAHbubPOkV2kmXKLir1Bg==","signed_message":"bundle_sha256_bytes","signed_at":"2026-08-06T02:36:36.372929Z","bundle_sha256":"90354559e9207013695dbb6d55ddd96922a019ae689361f8005c5bd9a03c4ea9"}}