A Simple Baseline for Streaming Video Understanding

Yujiao Shen, Shulin Tian, Jingkang Yang, Ziwei Liu
S-Lab, Nanyang Technological University
SimpleStream teaser

SimpleStream uses only the most recent N frames with an off-the-shelf VLM, yet matches or surpasses complex streaming systems with memory banks, retrieval, and compression modules.

Abstract

Recent streaming video understanding methods increasingly rely on complex memory mechanisms to handle long video streams. We challenge this trend with a simple finding: a sliding-window baseline that feeds only the most recent N frames to an off-the-shelf VLM already matches or surpasses published streaming models. We formalize this baseline as SimpleStream and evaluate it against 13 major offline and online video LLM baselines on OVO-Bench and StreamingBench. Despite its simplicity, SimpleStream delivers consistently strong performance. With only 4 recent frames, it reaches 67.7% average accuracy on OVO-Bench and 78.95% on StreamingBench. Controlled ablations further reveal a consistent perception-memory trade-off: adding more historical context can improve recall, but often weakens real-time perception. This suggests that stronger memory, retrieval, or compression modules should not be taken as evidence of progress unless they clearly outperform a strong recent-context baseline under the same protocol. We therefore argue that future streaming benchmarks should separate recent-scene perception from long-range memory, so that gains from added complexity can be evaluated more clearly.

Streaming Video Understanding Landscape

Streaming landscape

Existing streaming methods add External Memory, Retrieval, Compression, or Latent Memory modules. SimpleStream uses only the most recent 4 frames as context.

Key Findings

1. Simple baselines are strong

Achieves 67.7% on OVO-Bench and 80.6% on StreamingBench under real-time evaluation, surpassing all 13 compared methods.

2. Perception-memory trade-off

More access to history can help recall, but it does not reliably improve overall performance and often weakens present-scene perception.

3. Longer context is not always better

Window size ablation shows accuracy is non-monotonic, peaking at 4 frames (67.7%). Beyond that, further growth yields flat or declining scores, contrary to the "more frames = better" assumption.

4. Benchmark design amplifies this advantage

Current benchmarks overweight perception-oriented tracks. Methods that preserve clear recent visual evidence are doubly rewarded: they align with backbone strengths and with how scores are computed.

Main Results

Comparison on OVO-Bench and StreamingBench. "Avg." = mean of Real-Time and Backward category averages on OVO-Bench. "StreamingBench" = RTVU accuracy.

Model #Frames StreamingBench OVO RT Avg. OVO Bwd Avg. OVO Avg.
Offline Video LLMs
Qwen2.5-VL-7B 1 fps73.3159.944.752.28
LLaVA-OneVision-7B 3271.1264.043.753.85
Qwen2-VL-7B 6469.0460.748.654.62
Online / Streaming Video LLMs
VideoLLM-online-8B 2 fps35.9920.817.719.26
Flash-VStream-7B 1 fps23.2328.427.427.90
Dispider-7B 1 fps67.6354.636.145.35
TimeChat-Online-7B 1 fps75.2861.941.751.80
StreamForest-7B 1 fps77.2661.252.056.60
Streamo-7B 1 fps--66.046.156.05
HERMES-7B 1 fps79.4469.049.459.20
SimpleStream (Ours)
Qwen2.5-VL + 4f 478.4778.451.965.13
Qwen2.5-VL + 8f 879.1176.650.863.70
Qwen3-VL + 4f 478.9581.454.067.70
Qwen3-VL + 8f 878.8379.954.967.37

Perception-Memory Trade-off

SimpleStream variants occupy the upper-right quadrant (high perception + competitive memory), while published streaming methods cluster in the lower-left region.

Perception-Memory quadrant plot

Efficiency Pareto Frontier

SimpleStream with 4 frames is Pareto-optimal: highest accuracy at lowest latency among all compared configurations.

Pareto efficiency plot

Perceptual Fidelity ΔP

ΔP measures the perception cost of each method relative to a recency baseline. Every published streaming method shows negative ΔP, indicating perception degradation from added complexity.

Perceptual fidelity bar plot

BibTeX

@article{simplestream2026,
  title     = {A Simple Baseline for Streaming Video Understanding},
  author    = {Shen, Yujiao and Tian, Shulin and Yang, Jingkang and Liu, Ziwei},
  journal   = {arXiv preprint arXiv:2604.02317},
  year      = {2026},
}