Intel / post

T*: Turning Long-Video Understanding Into a Search Problem

Source: x.com
Date: Published Oct 22, 2025

Most models watch every frame - T* learns to search. In our latest blog post, we show how T* rethinks long-form video understanding as temporal search, finding the “needles” in long video haystacks with just a few key frames. https://t.co/CPmM49mCpE

Oct 22, 2025 View on X

Context

Understanding a long video with an AI model usually means processing many of its frames. Stanford AI Lab describes T*, which treats long-form video understanding as a search for the few key frames that answer a question, the needles in a video haystack, instead of watching every frame.

The paper frames this as reframing temporal search as spatial search, using adaptive zooming to decide which parts of the video to examine more closely. To test it, the authors built LV-Haystack, a dataset of 480 hours of video with 15,092 human-annotated instances. With a budget of 32 frames, the paper reports GPT-4o's accuracy on a long-video rising from 50.5% to 53.1%, and an open vision-'s from 56.5% to 62.4%.

Those are measured gains under a fixed frame budget on one benchmark setup, not a claim that T* solves long-video understanding.

Terms in this piece · Glossary

benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.

Read the source x.com