Novel or Not? PatEx-Gym: Agentic Prior-Art Search for Novelty Assessment in Patent Examination

Novel or Not? PatEx-Gym. Agentic Prior-Art Search for Novelty Assessment in Patent Examination. By Qiyao Wang and seven others, from the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, the University of Chinese Academy of Sciences, and two other institutions. Published on the twenty-fourth of October. Its code repository is worked out: about two hundred source files, with a readme.

Patent examination is a hard, real-world task for an agent. It has to discover prior art that is sparse, refine its queries over many rounds, inspect technical evidence, and produce a novelty assessment written up in an Office Action. All of this happens across heterogeneous scientific and technical domains. Existing benchmarks mostly work with static, limited sets of candidates, so the full process of search, reasoning and assessment has gone largely unexamined.

PatEx-Gym is presented as the first framework for developing and evaluating agents for patent examination. What it measures is the whole pipeline: not only whether the final decision on novelty is right, but whether the agent finds the prior art and documents its assessment in an Office Action. The agent is not handed a short list of candidates to rank. It has to go looking, in a very large collection, for the few documents that matter, and then reason from what it finds to a written judgment. Each of those stages can fail on its own, and the framework is designed so that each can be seen.

The framework has three components. The first is PatEx-Bench. It provides one thousand realistic examination tasks and two large retrieval corpora: one of more than a million patents, and one of fifty-two thousand scientific papers. A harness runtime supports iterative search, document visits, reasoning, and generation of the Office Action.

The second component is PASS, an annotation-free framework for synthesising prior-art agentic search trajectories. It treats document visits as implicit signals of relevance and as anchors for evidence acquisition. With those it models query behaviour within a turn, search progression across turns, and evidence and outcome quality across a whole trajectory. The output is a dataset called PatEx-SynData. Because no annotation is required, the supervision comes from the behaviour recorded in search itself: which documents an agent chose to open is taken as a sign of what was relevant.

The third component is PatEx-Agent, a four billion parameter model trained by supervised fine-tuning on the curated trajectories, intended as a lightweight baseline.

Extensive experiments yield two lessons about measurement. Final decision accuracy alone does not fully reflect the quality of an examination. And discovering prior art remains difficult, even for strong frontier models. Against substantially larger general-purpose agents, PatEx-Agent performs competitively. The authors conclude that training on PatEx-SynData lets a compact model of four billion parameters serve as a competitive baseline for prior-art search and patent examination.

A correct verdict on novelty can hide a poor search behind it, which is why the benchmark scores the search too. The finding that frontier models still struggle to discover prior art says the hard part is not the final judgment but getting to the evidence in the first place. And the small agent shows that the synthetic trajectories carry enough of that skill to be useful. The authors offer patent examination as a demanding testbed for agentic search and for deep research agents more broadly.