{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:QL3I3ZIGEJMOY3P5YGHF7IFB7E","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":"7425937d2c1e0b971a6502239a26ff698ee0bcd215c195eb1e6a98d6679ddda8","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-12T15:56:22Z","title_canon_sha256":"39a769729639f87ff4c5387af0d41e3bc3768ae72fd97f8ee86b78d025b233e2"},"schema_version":"1.0","source":{"id":"2605.12309","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.12309","created_at":"2026-05-20T00:00:43Z"},{"alias_kind":"arxiv_version","alias_value":"2605.12309v2","created_at":"2026-05-20T00:00:43Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.12309","created_at":"2026-05-20T00:00:43Z"},{"alias_kind":"pith_short_12","alias_value":"QL3I3ZIGEJMO","created_at":"2026-05-20T00:00:43Z"},{"alias_kind":"pith_short_16","alias_value":"QL3I3ZIGEJMOY3P5","created_at":"2026-05-20T00:00:43Z"},{"alias_kind":"pith_short_8","alias_value":"QL3I3ZIG","created_at":"2026-05-20T00:00:43Z"}],"graph_snapshots":[{"event_id":"sha256:353ede5a143586a711ca7c2f674ebbaccb02e31d984bf32b51f69886659b1d60","target":"graph","created_at":"2026-05-20T00:00:43Z","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":"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."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","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."}],"snapshot_sha256":"815f6c2fbd94908d792d62590e99c63d628330ec8cf270d6c8c8d22fcabbd8ad"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[{"findings_count":0,"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"}],"endpoint":"/pith/2605.12309/integrity.json","findings":[],"snapshot_sha256":"3941e578d5f4b970e0ed41b0bf3fdfdad87bb66ff5fb1c30dca739f08b3a0b74","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"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","authors_text":"Junxian Li, Kai Liu, Renjing Pei, Yulun Zhang, Zhikai Chen, Zhixin Wang, Zizhong Ding","cross_cats":[],"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.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-12T15:56:22Z","title":"G$^2$TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models"},"references":{"count":44,"internal_anchors":11,"resolved_work":44,"sample":[{"cited_arxiv_id":"2303.08774","doi":"","is_internal_anchor":true,"ref_index":1,"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Improving image generation with better captions","work_id":"ea79c0c7-072c-44cd-8531-ee4a28fe2a2f","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"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","year":2024},{"cited_arxiv_id":"2501.17811","doi":"","is_internal_anchor":true,"ref_index":4,"title":"Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling","work_id":"67d9e391-26d1-459e-ab56-07e60511c886","year":2025},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Diffusion models in vision: A survey.TPAMI","work_id":"01288b4b-f476-46c0-a59f-3c6b857a75ad","year":2023}],"snapshot_sha256":"e64a5587a31f7acbbf214a7e0efc5670e3366a2eca512323528652e81fd61f16"},"source":{"id":"2605.12309","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-19T16:49:05.384592Z","id":"0da71e8b-4374-4707-a2d9-6e37374a51d1","model_set":{"reader":"grok-4.3"},"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","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.","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.","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."}},"verdict_id":"0da71e8b-4374-4707-a2d9-6e37374a51d1"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:81b9d32d4f4a5a67e679a80290a4e387a50a2fdaa74c1e27cbed4752f4193e65","target":"record","created_at":"2026-05-20T00:00:43Z","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":"7425937d2c1e0b971a6502239a26ff698ee0bcd215c195eb1e6a98d6679ddda8","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-12T15:56:22Z","title_canon_sha256":"39a769729639f87ff4c5387af0d41e3bc3768ae72fd97f8ee86b78d025b233e2"},"schema_version":"1.0","source":{"id":"2605.12309","kind":"arxiv","version":2}},"canonical_sha256":"82f68de5062258ec6dfdc18e5fa0a1f922d2716418558126e73a636cd567f0e9","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"82f68de5062258ec6dfdc18e5fa0a1f922d2716418558126e73a636cd567f0e9","first_computed_at":"2026-05-20T00:00:43.036862Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-20T00:00:43.036862Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"3G86btMA2y8XHgRF5BNEEnl5/n2begJpry/hzfV62XFwq+N+9x+xhhhob17FlAL0nLsfIJOPv+XMfuvg6qhIAg==","signature_status":"signed_v1","signed_at":"2026-05-20T00:00:43.037600Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.12309","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:81b9d32d4f4a5a67e679a80290a4e387a50a2fdaa74c1e27cbed4752f4193e65","sha256:353ede5a143586a711ca7c2f674ebbaccb02e31d984bf32b51f69886659b1d60"],"state_sha256":"e1b6fbb02c003226791be26a004e4e2f0a325ad52b5551951f6ec6a370648242"}