{"paper":{"title":"OmniSIFT: Modality-Asymmetric Token Compression for Efficient Omni-modal Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"OmniSIFT reduces omni-modal token sequences to 25 percent of their length while matching or exceeding full-context accuracy on several tasks.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Bohan Zeng, Bozhou Li, Jiaheng Liu, Junfei Wu, Jungang Li, Liang Wang, Pengfei Wan, Qiang Liu, Xinlong Chen, Xuyang Liu, Yang Shi, Yiyan Ji, Yuanxing Zhang, Yue Ding, Yushuo Guan","submitted_at":"2026-02-04T17:51:05Z","abstract_excerpt":"Omni-modal Large Language Models (Omni-LLMs) have demonstrated strong capabilities in audio-video understanding tasks. However, their reliance on long multimodal token sequences leads to substantial computational overhead. Despite this challenge, token compression methods designed for Omni-LLMs remain limited. To bridge this gap, we propose OmniSIFT (Omni-modal Spatio-temporal Informed Fine-grained Token compression), a modality-asymmetric token compression framework tailored for Omni-LLMs. Specifically, OmniSIFT adopts a two-stage compression strategy: (i) a spatio-temporal video pruning modu"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"With merely 25% of the original token context, OmniSIFT consistently outperforms all compression baselines and even surpasses the performance of the full-token model on several tasks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the spatio-temporal pruning and vision-guided audio selection preserve all task-critical information across diverse video-audio content without introducing systematic bias.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"OmniSIFT delivers modality-asymmetric token compression for omni-LLMs via spatio-temporal video pruning and vision-guided audio selection, outperforming baselines at 25% token count.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"OmniSIFT reduces omni-modal token sequences to 25 percent of their length while matching or exceeding full-context accuracy on several tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f84376d72f1a270cb1af84fc693fa689699909ab1be784bb1ed8f4afd857eea7"},"source":{"id":"2602.04804","kind":"arxiv","version":2},"verdict":{"id":"1c2b82ee-18a6-486d-8527-6019016f34fe","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T07:21:07.168044Z","strongest_claim":"With merely 25% of the original token context, OmniSIFT consistently outperforms all compression baselines and even surpasses the performance of the full-token model on several tasks.","one_line_summary":"OmniSIFT delivers modality-asymmetric token compression for omni-LLMs via spatio-temporal video pruning and vision-guided audio selection, outperforming baselines at 25% token count.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the spatio-temporal pruning and vision-guided audio selection preserve all task-critical information across diverse video-audio content without introducing systematic bias.","pith_extraction_headline":"OmniSIFT reduces omni-modal token sequences to 25 percent of their length while matching or exceeding full-context accuracy on several tasks."},"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"}