New AI Method Helps Chatbots Finally Read Long PDFs Correctly

A team of researchers has published a new method called PILAR that makes AI assistants noticeably better at a task most of us take for granted: reading a long document and pulling out the right answer. Think of the 200-page financial report, the insurance policy, or the scientific paper where the answer you need is split across a table on page 3, a chart on page 40, and a paragraph on page 150. Today's AI tools usually read those chunks one at a time and lose the thread. PILAR stitches them together into a single connected map of facts before answering.

The idea is simple to picture. Imagine a librarian who, instead of handing you one book at a time, first builds an index linking every name, number, and claim across the entire shelf — then answers your question using that index. That is essentially what PILAR does. It treats sentences, tables, and figures as pieces of evidence and links the ones that refer to the same people, places, or ideas. When you ask a question, the AI can follow those links rather than just grabbing the closest page.

The results are modest but real. Tested across four different AI agent setups (AI that can take actions, like searching and reading on its own), fourteen search systems, and two standard question sets, PILAR came out on top. Its biggest wins were on questions requiring multiple steps — for example, "Which subsidiary of the company named in the 2019 filing had the highest revenue in 2022?" — improving accuracy by 5.9 points on the hardest three-step questions, and 2.9 points on questions requiring two facts to be combined. That is the difference between a confident wrong answer and a correct one.

The honest catch: the researchers found that nearly all the gains came from text. Charts and figures only helped once the system was carefully filtered to look at the right region of the page. In other words, AI still struggles to truly 'see' a graph the way a human does. So PILAR is a step forward for document-reading AI, not a finished product. It was accepted to a peer-reviewed conference and is aimed at researchers, but the underlying trick — linking evidence instead of just retrieving pages — is likely to show up in commercial AI assistants before long.