AI Paper Atlas

Research map / Video / Video-language understanding

video-SALMONN S: Memory-Enhanced Streaming Audio-Visual LLM

Guangzhi Sun, Yixuan Li, Xiaodong Wu, Yudong Yang, Wei Li, Zejun MA, Chao Zhang

ICML 2026 · Video › Video-language understanding · approach: Long video understanding

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace similarity merging with TTT_MEM memory + discardingstreaming video LLM · ELViMaccuracy (%): 38.5 -> 43.9Improved
add long-span prediction loss to TTT-video (ablation)TTT · ELViM / LVBenchaccuracy (%): 42.3 -> 43.6 (ELViM)Improved
add TTT-video over Mamba-2 as memory (ablation)similarity discarding · ELViMaccuracy (%): 39.6 -> 42.3Improved
use K-means clustering instead of discarding/merging (ablation)token downsampling · ELViMaccuracy (%): 36.8 -> 33.8Hurt
add stage 2 scale-up training (ablation)TTT_MEM · long video benchmarksaccuracy (%): 43.6 -> 43.9 (ELViM)Improved
replace AdaReTaKe with modality-aware memory reading (ablation)video-SALMONN S · VideoEvalProaccuracy (%): 56.5 -> 58.9Improved
add TTT_MEM to all 4 deepstack streams (ablation)video-SALMONN S · LVBenchaccuracy (%): 55.4 -> 54.7Hurt
increase memory tokens from 16k to 128k (ablation)video-SALMONN S (w reading) · ELViMaccuracy (%): 46.7 -> 54.9Improved
increase memory reading at 128k memory (ablation)video-SALMONN S · Video-MMEaccuracy (%): 77.4 -> 77.3No clear effect
change long-span prediction span T from 2 to 1 or 32 (ablation)TTT_MEM · Video-MME longaccuracy (%): 71.3 -> 70.6 (T=1)Hurt
use video-SALMONN S instead of non-streaming Qwen3-VL on VideoEvalProQwen3-VL 8B · VideoEvalProaccuracy (%): 54.9 -> 58.9Improved
use PEMF or similarity merging versus non-streaming baselinestreaming baselines · VideoEvalProaccuracy (%): 54.9 -> 51.8Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy