KAIST's ReMo: training-free cross-modal token compression trims Omni-LLM inference cost
kaist-ai · hf · 2026-11-19
KAIST with University of Oxford present ReMo, a training-free framework that cuts input token cost for omni-modal LLMs (Omni-LLMs) at inference time.
- Motivation: Omni-LLMs reason jointly over audio, video, and text, but visual tokens dominate the input and are highly redundant.
- Method: Compress redundant visual tokens via cross-modal alignment in a shared space—a visual token is kept only if its information appears nowhere else; discarded object-level tokens are replaced with compact text proxies, redistributing visual information across audio and text streams.
- Results: Comparable accuracy on audio-visual tasks with far fewer tokens. Code is open-sourced.
More from Multimodal
- Scenario open-sources a full AI art team as a Photoshop plugin, MCP-ready — aziz4ai · 2026-11-19
- MAVE reconstructs editable 3D blockouts from film shots, works inside Blender — bdsqlsz · 2026-11-19
- AI-made Dragon Ball parody music video nails the anime look — Top_Brilliant_9587 · 2026-11-19
- GTA 6-style LoRA generates strikingly authentic rain-soaked neon artwork — freshstart2027 · 2026-11-19
- VAE Rhythm: Detecting AI Videos via Periodic 3D Causal VAE Decoding Traces — showlab · 2026-11-19
- NUS's Jeveration turns a frozen vision-language model into a single-pass image generator — NUS-DSA3101 · 2026-11-19