Patent review gets a trainable agent testbed: a 4B model holds its own against far larger general agents at prior-art search
Synopsis
Patent examination asks an agent to find sparse prior art, refine queries, inspect evidence and write a novelty judgment, yet existing benchmarks mostly stop at static candidate sets. PatEx-Gym turns that process into a runnable environment: PatEx-Bench supplies 1K examination tasks, retrieval corpora of over 1M patents and 52K scientific papers, and a runtime supporting iterative search and Office Action generation. PASS synthesizes search trajectories by treating document visits as implicit relevance signals, and the 4B PatEx-Agent fine-tuned on them performs competitively against substantially larger general-purpose agents.
Interpretation
Novelty assessment in patent examination becomes a trainable, evaluable agent task in which the agent must locate sparse prior art in a retrieval corpus, rewrite queries, visit documents and produce an Office Action. Earlier benchmarks largely offered static and limited candidate sets, leaving the combined search, reasoning and assessment process uncovered. PatEx-Bench provides 1K realistic examination tasks together with two large-scale retrieval corpora of over 1M patents and 52K scientific papers, plus a runtime supporting iterative search, document visits, reasoning and Office Action generation.
The training data needs no manual annotation; which documents the agent visited serve as relevance cues for synthesizing complete search trajectories. PASS uses document visits both as implicit relevance signals and as evidence-acquisition anchors, modeling intra-turn query behavior, inter-turn search progression, and trajectory-level evidence and outcome quality. The resulting PatEx-SynData is used for supervised fine-tuning of a 4B-scale model, yielding the lightweight baseline PatEx-Agent.
Experiments show final decision accuracy alone does not fully reflect examination quality, and prior-art discovery remains difficult even for strong frontier models. In this difficult setting, the 4B PatEx-Agent achieves competitive performance against substantially larger general-purpose agents. The comparison comes from extensive experiments on PatEx-Gym, indicating that training on PatEx-SynData lets a compact model serve as a usable baseline for prior-art search and patent examination.
Perspective
The demonstrated scope is the PatEx-Gym framework itself: 1K examination tasks, two large-scale retrieval corpora, a runtime supporting iterative search and Office Action generation, and a 4B model fine-tuned on PatEx-SynData. The intended readers are researchers and engineers working on agentic retrieval and deep-research capabilities.
The material notes that final decision accuracy alone cannot fully reflect examination quality and that prior-art discovery remains difficult even for strong frontier models; these conclusions are reported at summary level, and task composition and evaluation details are not expanded in the material.