AIRA_2: Overcoming Bottlenecks in AI Research Agents

AIRA_2 addresses three critical bottlenecks in AI research agents: synchronous single-GPU execution limiting throughput, generalization gaps in validation-based selection, and fixed LLM operator capacity constraints. The solution introduces asynchronous multi-GPU processing, Hidden Consistent Evaluation protocol, and ReAct agents with dynamic action scoping. On MLE-bench-30, AIRA_2 achieves 71.8% percentile rank in 24 hours, improving to 76.0% at 72 hours, surpassing previous best results of 69.