AIRA_2：克服人工智能研究代理中的瓶颈

AIRA$_2$ addresses three structural performance bottlenecks in AI research agents. First, synchronous single-GPU execution constrains sample throughput, limiting the benefits of search. Second, a generalization gap arises where validation-based selection leads to performance degradation over extended search horizons. Third, the limited capability of fixed, single-turn large language model (LLM) operators imposes a ceiling on search performance. To overcome these challenges, AIRA$_2$ introduces three architectural choices: an asynchronous multi-GPU worker pool that increases experiment throughput linearly; a Hidden Consistent Evaluation protocol that delivers a reliable evaluation signal; and ReAct agents that dynamically scope their actions and debug. These innovations significantly enhance the efficiency and effectiveness of AI research agents.