deepexperience/hypereyes

HyperEyes is a research project focused on making AI agents smarter and faster at answering questions that require searching the internet while also looking at images. When you ask an AI agent a complex question, it often has to search for information multiple times in sequence: look up one thing, get the result, then look up another thing, and so on. This is slow and wasteful when many of those searches are independent and could happen at the same time. HyperEyes introduces a new approach it calls search wider, not longer. Instead of chaining searches one after another, the system identifies all the things it needs to look up and launches those searches simultaneously in a single round. It extends this parallel approach to visual content too: it can identify multiple objects or people in an image and search for information about all of them at the same time, rather than one by one. To train the AI to actually prefer this efficient behavior, the researchers developed a two part reinforcement learning system. One part rewards the agent for completing tasks with fewer tool call rounds overall. The other part provides detailed correction signals when the agent makes mistakes, helping it learn even when the training signal is sparse. The project also introduces a new benchmark called IMEB, a set of 300 human curated test questions that measure both accuracy and efficiency together, making it possible to compare agents fairly on whether they get the right answer and how many searches it took them to get there. The code is written in Python and is accompanied by a published research paper.