Blind Trust Problem Causes 15-30 Point Drops in Frontier Video AI

Video reasoning models have a structural blind spot that gets worse when the camera shakes or the sun hits the lens wrong. A paper published on arXiv by Yangfan He, Yujin Choi, and Jaehong Yoon identifies what they call the "Blind Trust Problem": current frontier models treat every video frame as equally reliable, with no mechanism to discount corrupted or low-quality frames during reasoning. Under motion blur, glare, and occlusion, that assumption causes frontier models to suffer accuracy drops of 15 to 30 percentage points.

The numbers are notable because they come from realistic visual corruptions, not adversarial attacks. Gemini-2.5-Pro scores 46.2% accuracy in the researchers' evaluation. Against that, the paper's proposed framework, Robust-TO, achieves 56.4% average accuracy on clean inputs and maintains 54.3% under five corruption types, recording the smallest clean-to-corrupted accuracy gap among compared methods and 5.8 percentage points above the strongest open-source baseline under corruption.

Robust-TO addresses the Blind Trust Problem through per-frame trustworthiness integration: rather than treating the video stream as uniformly credible, the framework assigns calibrated reliability scores to each tool output and applies a three-tier synthesis process weighting high, medium, and low confidence signals at reasoning time. A confidence-cost GRPO reward trains the model to optimize correctness, reliability, and efficiency jointly.

The honest caveat is that this is a preprint, and whether the 15-30 point degradation holds across deployment contexts beyond the embodied benchmarks tested here is an open question. The paper evaluates five specific corruption types and does not address others such as compression artifacts or low-light conditions, so the robustness picture may be incomplete. What the inference overhead of per-frame scoring looks like at production frame rates is also not addressed.

For teams building video AI into environments where clean footage is not guaranteed, the failure mode the paper names is real and the direction worth watching. A system that degrades gracefully under visual noise is a different product from one that performs on clean benchmarks only, and the Blind Trust Problem now has a name and a benchmark attached to it.