Cardano — Learning When to Stop Watching

Overview

A streaming video model normally watches a clip to the end before answering. Cardano decides, while the video is still playing, that watching more of it would not change its answer — and stops there.

The decision is made by a 40-parameter linear head that reads nine features of the model's own answer distribution. It reads no hidden state, runs at a single global operating point, and is never tuned per dataset. The same head, fit once, transfers unchanged across three backbones and onto benchmarks it was never fit on.

Cardano pipeline: windows arrive, a KV cache accumulates, and a commit head reads the answer distribution after each window. — Video arrives one window at a time. The backbone keeps a key–value cache, so each frame is encoded once and windows past the commit are never encoded at all. After each window the head asks a single question: *would the rest of the video change this answer?*

A stopping rule is not a confidence threshold. Confidence asks how sure am I?; Cardano asks would more evidence move me? Those come apart — and the practical consequence is that Cardano matches a tuned per-dataset confidence threshold while itself being untuned.

Results

Accuracy and fraction of video watched on Qwen2.5-VL-7B, against a matched causal budget that watches the same number of leading windows:

Benchmark	Cardano	Watched	vs. matched budget
STAR	67.2	0.69	+3.7 [+2.4, +5.1]
NExT-QA	81.8	0.31	+1.3 [+0.3, +2.3]
FunQA	76.0	0.41	+1.9 [+0.9, +3.0]
Video-MME	65.1	0.59	+2.5 [+0.6, +4.4]

Every interval excludes zero and holds on 10 of 10 head-fit seeds. On LLaVA-OneVision-7B the same unrefit head gives +3.5, +2.3 and +2.4 on the first three.

The saving is what the stopping head buys. Where accuracy also improves on an off-the-shelf backbone, that comes from fine-tuning the answerer.

Accuracy against fraction of video watched, on three benchmarks. Cardano sits above the causal fixed-budget curve at every operating point. — Green points are the four budget settings. Cardano sits above the causal fixed-budget curve at every one, and on FunQA above the whole-video backbone while watching under half the video.

Supplementary — episodes

Eight episodes, one strip each. All sixteen windows are shown; the boxed window is where the head committed, and everything to its right is faded because the model never encoded it. Two episodes are failure cases, included deliberately.

commit window, answered correctly commit window, answered wrongly never encoded

The mechanism on safety-critical footage

A pedestrian is already crossing in the first window. The head commits at the third of sixteen and answers correctly, so the answer exists while the clip is still playing rather than after it ends.

A VRU-Accident dashcam episode. The commit window is boxed at window three of sixteen; the remaining thirteen windows are faded. — VRU-Accident. Thirteen of sixteen windows are never encoded.

The converse matters as much. On egocentric planning — where the question asks what to do next, so the decisive evidence is the final frame — the same head watches 0.89 of the stream and declines to commit early on about two thirds of items. That is the correct behaviour there, and it is not something a fixed budget can do.

BibTeX

@misc{galoaa2026cardano,
  title  = {Cardano: Learning When to Stop Watching},
  author = {Galoaa, Bishoy and Hasan, Sadid A. and Ostadabbas, Sarah},
  year   = {2026},
  note   = {Code: https://github.com/ostadabbas/Cardano;
            Weights: https://huggingface.co/bishoygaloaa/cardano}
}