New Trick Lets AI Read Huge Documents 10 Times Faster

When you paste a 200-page contract or a long email thread into an AI chatbot, the model doesn't actually skim. It reads everything, twice over: once to take in your question and once to write its answer. That's why long documents are slow, expensive, and often cut off with a message like "this conversation is too long." The new research tackles exactly that bottleneck.

The trick, called Self-Indexing Attention, is to give the AI a bookmark list. As it reads, the system quietly notes which words or chunks seem important using a single bit of information each — essentially a yes-or-no flag, like highlighting a page instead of copying it. Later, when you ask a question, the AI checks the flags first and only re-reads the flagged sections. Because the flags are just ones and zeros, the computer can sift through millions of them using a simple bit-by-bit comparison, a trick modern chips already do extremely quickly.

The results are striking. The system looked at only 5% of the text it normally would, yet scored almost as well as the full, careful-reading version on two standard tests for long-document understanding. Speed improved by up to 6.1 times while the AI was reading and 10.3 times while it was writing an answer. It also plays nicely with other popular memory-saving techniques, including TurboQuant and DeepSeekV4-Flash, and requires no extra training — you can bolt it onto existing models.

One honest caveat: this is a research paper, not a product you can use today. The tests are academic reading-comprehension benchmarks, not messy real-world tasks, and the speedups are measured on the specific part of the AI that decides what to look at, not the whole system. Still, if it holds up, the practical payoff is straightforward — cheaper, faster AI that can handle far longer documents before hitting a limit.