PatEx-Gym Trains AI Agents to Find Prior Art: 1M Patents and a 4B Model

Patent examination has long been considered one of the most awkward tasks for AI. The system has to find rare documents that undermine an application's novelty, and it has to do so among millions of texts from different technical fields. A team made up of Qiyao Wang, Xinyi Chen, Hongbo Wang, Jiayuan Lu, Jianhang Li, Minghuan Tan, Yuan Lin and Min Yang has introduced PatEx-Gym. It is the first environment in which an agent goes through the examiner's entire workflow: search, reading, reasoning and a final conclusion. The environment includes 1000 realistic tasks, a corpus of more than 1M patents and 52 thousand scientific papers. The reference agent, PatEx-Agent, has only 4B parameters.

Why Older Benchmarks Didn't Reflect an Examiner's Work

Most existing tests hand the model a ready-made, limited set of candidates. The task boils down to picking the right item from what is shown. A real examiner works differently. They formulate a query, look at the results, open documents, realize they were searching for the wrong thing, and reformulate the search. The result is written up as an Office Action, an official conclusion in which the finding on novelty must rest on specific sources.

PatEx-Gym carries this logic over into an interactive environment. The agent uses search tools, "visits" documents, reasons, and generates a conclusion at the end. The authors call the task long-horizon: an error at an early search step cannot later be fixed by a polished final text.

What PatEx-Gym Consists Of

The framework combines three components and rests on four artifacts:

• PatEx-Bench — a benchmark with 1000 realistic examiner tasks, two large search corpora (over 1M patents and 52K scientific papers) and an executable environment. The environment supports iterative search, document visits, reasoning and Office Action generation.
• PASS (Prior-Art Agentic Search Trajectory Synthesis) — a method for synthesizing training search trajectories without manual annotation. It uses document visits as an implicit relevance signal and as anchor points for gathering evidence. These signals are used to model three levels: query behavior within a single turn, how the search develops across turns, and the quality of evidence and outcome over the whole trajectory.
• PatEx-SynData — a dataset produced through PASS from curated trajectories.
• PatEx-Agent — a 4B-parameter agent trained with supervised fine-tuning on these trajectories. The authors position it as a lightweight baseline.

The PASS approach shows how to avoid expensive manual annotation. Instead of asking experts to describe the correct search path, the method takes the bare fact of which documents the agent opened and treats it as a sign that they were useful.

What the Experiments Showed

The main finding concerns metrics: the final accuracy of the "novel or not" decision does not reflect the quality of the examination. A model can reach the correct verdict without finding the key documents, and from a legal standpoint such a conclusion is poorly substantiated. For that reason, PatEx-Bench evaluates the process itself, including the discovery of prior art.

This part turned out to be the hardest. Finding rare relevant documents remains a problem even for strong frontier models. At the same time, the authors state that PatEx-Agent 4B, after training on PatEx-SynData, delivers results competitive with significantly larger general-purpose agents. For a narrow professional task, this is an argument in favor of specialized small models.

The authors see patent examination as a demanding testbed for agentic search and deep research agents. The task combines long chains of actions, ambiguous relevance and the need to document conclusions.

Who Can Use This Right Now

The code is published in the AIforIP/PatEx-Gym repository, and there is also a project page. Teams building deep research agents or search systems for technical literature get a ready-made testbed with a thousand tasks and a large corpus. Patent offices and law firms can use it to compare tools not only by verdict but also by the evidence they find. For model developers, the PASS approach itself is valuable: it shows how to build training data for agents from behavior logs without manual annotation.