{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:QL3I3ZIGEJMOY3P5YGHF7IFB7E","short_pith_number":"pith:QL3I3ZIG","schema_version":"1.0","canonical_sha256":"82f68de5062258ec6dfdc18e5fa0a1f922d2716418558126e73a636cd567f0e9","source":{"kind":"arxiv","id":"2605.12309","version":2},"attestation_state":"computed","paper":{"title":"G$^2$TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Generation-guided selection from the VAE latent cuts visual tokens by 1.94x in separate-encoder unified multimodal models while preserving both reasoning accuracy and editing quality.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Junxian Li, Kai Liu, Renjing Pei, Yulun Zhang, Zhikai Chen, Zhixin Wang, Zizhong Ding","submitted_at":"2026-05-12T15:56:22Z","abstract_excerpt":"The development of separate-encoder Unified multimodal models (UMMs) comes with a rapidly growing inference cost due to dense visual token processing. In this paper, we focus on understanding-side visual token reduction for improving the efficiency of separate-encoder UMMs. While this topic has been widely studied for MLLMs, existing methods typically rely on attention scores, text-image similarity and so on, implicitly assuming that the final objective is discriminative reasoning. This assumption does not hold for UMMs, where understanding-side visual tokens must also preserve the model's cap"},"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":"2605.12309","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-12T15:56:22Z","cross_cats_sorted":[],"title_canon_sha256":"39a769729639f87ff4c5387af0d41e3bc3768ae72fd97f8ee86b78d025b233e2","abstract_canon_sha256":"7425937d2c1e0b971a6502239a26ff698ee0bcd215c195eb1e6a98d6679ddda8"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:00:43.037600Z","signature_b64":"3G86btMA2y8XHgRF5BNEEnl5/n2begJpry/hzfV62XFwq+N+9x+xhhhob17FlAL0nLsfIJOPv+XMfuvg6qhIAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"82f68de5062258ec6dfdc18e5fa0a1f922d2716418558126e73a636cd567f0e9","last_reissued_at":"2026-05-20T00:00:43.036862Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:00:43.036862Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"G$^2$TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Generation-guided selection from the VAE latent cuts visual tokens by 1.94x in separate-encoder unified multimodal models while preserving both reasoning accuracy and editing quality.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Junxian Li, Kai Liu, Renjing Pei, Yulun Zhang, Zhikai Chen, Zhixin Wang, Zizhong Ding","submitted_at":"2026-05-12T15:56:22Z","abstract_excerpt":"The development of separate-encoder Unified multimodal models (UMMs) comes with a rapidly growing inference cost due to dense visual token processing. In this paper, we focus on understanding-side visual token reduction for improving the efficiency of separate-encoder UMMs. While this topic has been widely studied for MLLMs, existing methods typically rely on attention scores, text-image similarity and so on, implicitly assuming that the final objective is discriminative reasoning. This assumption does not hold for UMMs, where understanding-side visual tokens must also preserve the model's cap"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Experiments on image understanding and editing benchmarks show that G²TR substantially reduces visual tokens and prefill computation by 1.94x while maintaining both reasoning accuracy and editing quality, outperforming baselines on almost all benchmarks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That token importance estimated from consistency with VAE latent provides a task-agnostic signal that preserves the model's editing and generation capabilities without degradation, even though the selection is performed only on the understanding-side tokens.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"G²TR reduces visual tokens and prefill compute by 1.94x in separate-encoder UMMs via generation-guided importance from VAE latent consistency, balanced selection, and merging, while preserving reasoning accuracy and editing quality.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Generation-guided selection from the VAE latent cuts visual tokens by 1.94x in separate-encoder unified multimodal models while preserving both reasoning accuracy and editing quality.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"815f6c2fbd94908d792d62590e99c63d628330ec8cf270d6c8c8d22fcabbd8ad"},"source":{"id":"2605.12309","kind":"arxiv","version":2},"verdict":{"id":"0da71e8b-4374-4707-a2d9-6e37374a51d1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T16:49:05.384592Z","strongest_claim":"Experiments on image understanding and editing benchmarks show that G²TR substantially reduces visual tokens and prefill computation by 1.94x while maintaining both reasoning accuracy and editing quality, outperforming baselines on almost all benchmarks.","one_line_summary":"G²TR reduces visual tokens and prefill compute by 1.94x in separate-encoder UMMs via generation-guided importance from VAE latent consistency, balanced selection, and merging, while preserving reasoning accuracy and editing quality.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That token importance estimated from consistency with VAE latent provides a task-agnostic signal that preserves the model's editing and generation capabilities without degradation, even though the selection is performed only on the understanding-side tokens.","pith_extraction_headline":"Generation-guided selection from the VAE latent cuts visual tokens by 1.94x in separate-encoder unified multimodal models while preserving both reasoning accuracy and editing quality."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.12309/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-19T22:41:58.286181Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T10:39:20.227335Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T08:31:16.589132Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T07:40:35.565880Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"3941e578d5f4b970e0ed41b0bf3fdfdad87bb66ff5fb1c30dca739f08b3a0b74"},"references":{"count":44,"sample":[{"doi":"","year":2023,"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","ref_index":1,"cited_arxiv_id":"2303.08774","is_internal_anchor":true},{"doi":"","year":2023,"title":"Improving image generation with better captions","work_id":"ea79c0c7-072c-44cd-8531-ee4a28fe2a2f","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models","work_id":"f5f8f484-9d8c-4b18-bcf2-5b407818c91e","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling","work_id":"67d9e391-26d1-459e-ab56-07e60511c886","ref_index":4,"cited_arxiv_id":"2501.17811","is_internal_anchor":true},{"doi":"","year":2023,"title":"Diffusion models in vision: A survey.TPAMI","work_id":"01288b4b-f476-46c0-a59f-3c6b857a75ad","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":44,"snapshot_sha256":"e64a5587a31f7acbbf214a7e0efc5670e3366a2eca512323528652e81fd61f16","internal_anchors":11},"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":"2605.12309","created_at":"2026-05-20T00:00:43.036967+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.12309v2","created_at":"2026-05-20T00:00:43.036967+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.12309","created_at":"2026-05-20T00:00:43.036967+00:00"},{"alias_kind":"pith_short_12","alias_value":"QL3I3ZIGEJMO","created_at":"2026-05-20T00:00:43.036967+00:00"},{"alias_kind":"pith_short_16","alias_value":"QL3I3ZIGEJMOY3P5","created_at":"2026-05-20T00:00:43.036967+00:00"},{"alias_kind":"pith_short_8","alias_value":"QL3I3ZIG","created_at":"2026-05-20T00:00:43.036967+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/QL3I3ZIGEJMOY3P5YGHF7IFB7E","json":"https://pith.science/pith/QL3I3ZIGEJMOY3P5YGHF7IFB7E.json","graph_json":"https://pith.science/api/pith-number/QL3I3ZIGEJMOY3P5YGHF7IFB7E/graph.json","events_json":"https://pith.science/api/pith-number/QL3I3ZIGEJMOY3P5YGHF7IFB7E/events.json","paper":"https://pith.science/paper/QL3I3ZIG"},"agent_actions":{"view_html":"https://pith.science/pith/QL3I3ZIGEJMOY3P5YGHF7IFB7E","download_json":"https://pith.science/pith/QL3I3ZIGEJMOY3P5YGHF7IFB7E.json","view_paper":"https://pith.science/paper/QL3I3ZIG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.12309&json=true","fetch_graph":"https://pith.science/api/pith-number/QL3I3ZIGEJMOY3P5YGHF7IFB7E/graph.json","fetch_events":"https://pith.science/api/pith-number/QL3I3ZIGEJMOY3P5YGHF7IFB7E/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/QL3I3ZIGEJMOY3P5YGHF7IFB7E/action/timestamp_anchor","attest_storage":"https://pith.science/pith/QL3I3ZIGEJMOY3P5YGHF7IFB7E/action/storage_attestation","attest_author":"https://pith.science/pith/QL3I3ZIGEJMOY3P5YGHF7IFB7E/action/author_attestation","sign_citation":"https://pith.science/pith/QL3I3ZIGEJMOY3P5YGHF7IFB7E/action/citation_signature","submit_replication":"https://pith.science/pith/QL3I3ZIGEJMOY3P5YGHF7IFB7E/action/replication_record"}},"created_at":"2026-05-20T00:00:43.036967+00:00","updated_at":"2026-05-20T00:00:43.036967+00:00"}