The Patent Examination Gym uses agentic prior-art search to determine whether an invention is novel.
PatEx-Gym is an agentic benchmark for patent novelty, testing prior-art search, evidence inspection, strategy revision, and Office Actions.
Key Takeaways
- PatEx-Gym is a new patent AI benchmark described as the first long-horizon agentic task in the patent domain.
- It evaluates patent agents across more than 1 million patents and 52,000 scientific papers in an interactive environment.
- The benchmark tests the full workflow, including prior-art search, reasoning, evidence inspection, and writing grounded Office Actions.
- Results can reveal process-level failures, such as missing relevant prior art, ignoring technical details, or failing to connect evidence to conclusions.
Meta description: PatEx-Gym tests patent agents over 1 million patents and 52,000 scientific papers, evaluating search, reasoning, evidence inspection, and Office Actions.
Tags: PatEx-Gym, patent AI benchmark, prior art search, agentic evaluation, Office Actions

What PatEx-Gym tests
For readers building or evaluating patent AI agents, PatEx-Gym is a new benchmark that describes itself as the first long-horizon agentic task in the patent domain. It moves beyond static documents and predefined candidate pools and tests whether an agent can search for prior art, inspect technical evidence, revise its search strategy, and produce evidence-grounded Office Actions. Prior art means existing patents or scientific literature that can show a claimed invention was already known. Office Actions are examination-style documents that identify objections and point to supporting evidence.
Why the full workflow matters
The evaluation covers the complete search–reasoning–assessment process inside an interactive tool environment spanning over 1 million patents and 52,000 scientific papers. A weak result may show more than a final-answer error: the agent may have missed relevant prior art, ignored technical details, or failed to connect evidence to a written conclusion.
Where to find the materials
For readers, the value is a clearer failure map. PatEx-Gym can help compare tools that do not just retrieve documents, but also decide what to retrieve next, how to interpret it, and how to write a grounded response. The paper is at https://arxiv.org/abs/2610.33503, the code is at https://github.com/AIforIP/PatEx-Gym, and the project page is at https://patex-gym.wangqiyao.me/. The project also asks for support and an upvote.
Comments (0)