PaSa is a paper search agent built by ByteDance that uses large language models to autonomously make decisions such as invoking search tools, reading papers, and selecting relevant references. The agent is optimized using reinforcement learning on a synthetic dataset of 35k academic queries and evaluated on a real-world benchmark. The project provides two 7B models (crawler and selector) on HuggingFace and significantly outperforms Google-based baselines and GPT-4o-powered approaches in recall and precision for scholarly search tasks.