Paper: A Simple Baseline for Streaming Video Understanding

Project: SimpleStream

Code: EvolvingLMMs-Lab/SimpleStream

Core Idea

SimpleStream is a deliberately simple streaming baseline: when a question arrives, keep only the most recent N observed frames and feed them with the question to an off-the-shelf VLM.

It uses:

  • 1 fps sampling under the observed-only streaming protocol;
  • a short recent-frame window, usually 2 / 4 / 8 frames;
  • Qwen2.5-VL-7B or Qwen3-VL-8B as the base VLM;
  • no new memory bank;
  • no retrieval module;
  • no KV cache compression;
  • no extra training.

So this is not a new architecture. It is a strong recency baseline for asking whether complex streaming memory is actually helping.

Benchmark

The paper evaluates on OVO-Bench and StreamingBench under a causal protocol: the model can only use frames observed before the query time. Numbers below follow Table 1 in the arXiv version.

MethodFramesStreamingBench RTVUOVO Real-TimeOVO BackwardOVO Avg.
StreamForest-7B1 fps77.2661.252.056.60
HERMES-7B1 fps79.4469.049.459.20
Qwen2.5-VL-7B + SimpleStream478.4778.451.965.13
Qwen2.5-VL-7B + SimpleStream879.1176.650.863.70
Qwen3-VL-8B + SimpleStream278.3179.353.566.38
Qwen3-VL-8B + SimpleStream480.5981.454.067.70
Qwen3-VL-8B + SimpleStream878.8379.954.967.37

The key result is simple: Qwen3-VL-8B + 4 recent frames reaches 67.70 OVO Avg. and 80.59 StreamingBench RTVU, beating the compared streaming baselines in the paper.

Takeaways

The main message is not that four frames solve long-video understanding. It is that current streaming benchmarks can reward clean recent-scene perception very strongly.

Two observations matter:

  • longer context is not monotonic: 8 frames is not always better than 4 frames;
  • extra historical retrieval can improve some memory tracks, but it may hurt real-time perception.

So future streaming VLM papers should compare against a strong recent-window baseline under the same backbone and protocol. Otherwise, it is hard to tell whether a memory module improves long-range reasoning or just loses less than a weak baseline.