Lookahead-R Optimizes Tool Retrieval Using Execution-Aware Surrogate World Models

Lookahead-R mitigates the latency-precision trade-off in agent tool retrieval by using a lightweight surrogate world model to predict execution success and cost. It employs uncertainty-guided Monte Carlo Tree Search to select tools under strict budget constraints, outperforming semantic retrievers on the ToolBench benchmark.