Model-driven arXiv retrieval workflow for building a paper set with a manual language parameter: initialize a run, fetch metadata for each model-designed query, let the model filter irrelevant items per query by keep indexes, then merge and dedupe into per-paper metadata directories. Use when query planning and relevance filtering should be done by the model, not rule-based heuristics.

基于模型的arXiv检索工作流，用于构建带手动语言参数的论文集：初始化运行，为每个模型设计的查询获取元数据，由模型根据保留索引过滤每条查询中的无关项，然后合并并去重，生成按论文划分的元数据目录。适用于查询规划和相关性过滤需由模型完成，而非依赖规则启发式方法的场景。