mohammad2012191/gridprobe

GridProbe is a research tool for making AI models that understand long videos more efficient when answering questions. These models, called Vision-Language Models or VLMs, normally look at a fixed set of video frames to answer a question, which can be wasteful if the answer depends on only a few key moments. GridProbe changes this by using the AI model's own internal confidence signals to identify which frames actually matter for a specific question, then directing the model to focus only on those frames. The core idea is a two-stage process. In the first stage, GridProbe samples a grid of candidate frames from the video and runs lightweight probes to build an importance map, a frame-by-frame picture of how relevant each moment is to the question being asked. In the second stage, the shape of that importance map determines how many frames to analyze in depth. Simple questions with clear evidence get fewer frames, harder questions with scattered evidence get more. This adaptive budget can reduce compute by several times compared to a fixed-frame approach. The tool works with Qwen3-VL models of different sizes (2B, 4B, and 8B parameters), requires no fine-tuning or additional training, and was evaluated on the Video-MME-v2 and LongVideoBench benchmarks. It is written in Python and requires a CUDA-compatible GPU (an A100 or better is recommended) along with Python 3.10 or later. It accompanies an academic paper released on arXiv in May 2026.