{"paper":{"title":"OmniDrop: Layer-wise Token Pruning for Omni-modal LLMs via Query-Guidance","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Layer-wise token pruning inside the LLM decoder, guided by text queries, allows omni-modal models to process audiovisual inputs faster while maintaining or improving accuracy.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Hyemi Jang, Jongsun Lee, Jooyoung Choi, Minseo Choi, Yeo Jeong Park, Yongkweon Jeon","submitted_at":"2026-05-14T06:54:37Z","abstract_excerpt":"Omni-modal large language models have demonstrated remarkable potential in holistic multimodal understanding; however, the token explosion caused by high-resolution audio and video inputs remains a critical bottleneck for real-time applications and long-form reasoning. Existing omni-modal token compression methods typically prune tokens at the input embedding level, relying on audio-video similarity or temporal co-occurrence as proxies for semantic relevance. In practice, such assumptions are often unreliable. To address this limitation, we propose OmniDrop, a training-free, layer-wise token p"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Experimental results across various audiovisual benchmarks demonstrate that OmniDrop outperforms all baselines by up to 3.58 points while reducing prefill latency by up to 40% and memory usage by up to 14.7%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That performing initial fusion in early layers followed by aggressive pruning in deeper layers, guided by text queries, reliably preserves semantic information without task-specific degradation.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"OmniDrop is a training-free layer-wise token pruning framework for omni-modal LLMs that uses query guidance and temporal diversity to reduce prefill latency by up to 40% and memory by 14.7% while improving benchmark scores by up to 3.58 points.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Layer-wise token pruning inside the LLM decoder, guided by text queries, allows omni-modal models to process audiovisual inputs faster while maintaining or improving accuracy.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"8068274d93d673443d4ac5aa68336733ebd366d5454111653e01eb7898750233"},"source":{"id":"2605.14458","kind":"arxiv","version":1},"verdict":{"id":"f678d1cc-701a-46a3-875e-9bf574146f08","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T01:32:16.886455Z","strongest_claim":"Experimental results across various audiovisual benchmarks demonstrate that OmniDrop outperforms all baselines by up to 3.58 points while reducing prefill latency by up to 40% and memory usage by up to 14.7%.","one_line_summary":"OmniDrop is a training-free layer-wise token pruning framework for omni-modal LLMs that uses query guidance and temporal diversity to reduce prefill latency by up to 40% and memory by 14.7% while improving benchmark scores by up to 3.58 points.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That performing initial fusion in early layers followed by aggressive pruning in deeper layers, guided by text queries, reliably preserves semantic information without task-specific degradation.","pith_extraction_headline":"Layer-wise token pruning inside the LLM decoder, guided by text queries, allows omni-modal models to process audiovisual inputs faster while maintaining or improving accuracy."},"references":{"count":33,"sample":[{"doi":"","year":2025,"title":"Divprune: Diversity-based visual token pruning for large multimodal models, 2025","work_id":"7d34fc13-2b41-48af-93a2-a4d706d22a94","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Token merging: Your vit but faster, 2023","work_id":"af1ae900-f884-4cb0-bebd-85d3b20d6230","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":"2c834314-7394-40d4-9cda-458a1172cdf8","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities","work_id":"008df105-2fdd-45d8-857a-8e35868aecb6","ref_index":4,"cited_arxiv_id":"2507.06261","is_internal_anchor":true},{"doi":"","year":2024,"title":"FlashAttention-2: Faster attention with better parallelism and work partitioning","work_id":"3c743fc7-8c9b-4a72-ad8a-26c5a7636140","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":33,"snapshot_sha256":"ac784e2b6b81fe29eccc8ca31a71cb2f97fb0e17b7aad919b3f8ed230aa9fc0a","internal_anchors":5},"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"}