Studying Without a Syllabus: Task-Agnostic Environment Preprocessing

Summary

This paper studies whether an LLM agent can explore an unfamiliar environment before test time, without a syllabus or knowledge of the downstream task distribution, and decide how to prepare it. The agent operates under a study budget and creates reusable artifacts, such as indices, scripts, or procedural guidance, for a frozen solver. The authors compare unaided and archive-equipped meta-agents with fixed synthetic-practice and corpus-processing methods across six heterogeneous benchmarks. A meta-agent achieves the highest Avg@3 reward on five of the six benchmarks, while fixed corpus processing performs best on the benchmark with the largest corpus. Increasing the study budget does not reliably improve downstream reward, so more preparation is not consistently better. However, the artifacts produced during study reduce the amount of test-time sampling required to reach a given score. The results suggest that environment preparation can move computation from repeated attempts during testing to a reusable pre-task study phase, although the best preparation strategy still depends on the environment.