Studying Without a Syllabus: Task-Agnostic Environment Preprocessing

Studying Without a Syllabus. Task-Agnostic Environment Preprocessing. By Vinay Samuel and five others, from Scale AI and the University of Maryland, College Park. Published on the ninth of September. It links to a code repository that is no longer there.

Before a language model agent tackles tasks in a new environment, it can look around first: inspect the available corpora and tools, and build reusable resources such as indices, scripts or procedural guidance. The question is whether it can do that without a syllabus, meaning before test time, with no knowledge of the tasks that will follow, and choosing for itself how to prepare. The answer here is a qualified yes. Preparing this way does not reliably raise the final reward as the study budget grows, but it does reduce the test-time sampling needed to reach a given score.

Most automated adaptation methods rely on task examples, trajectories or evaluation feedback to decide what to build. Existing task-agnostic approaches avoid that supervision, but commit in advance to one preparation strategy for a particular kind of environment. The authors instead formalize task-agnostic environment preprocessing, in which a studying system explores an environment under a budget and produces artifacts that a frozen solver then uses. The split matters. The solver does not change, so any difference in how it performs has to come from what the studying system left behind for it: an index, a script, a page of guidance.

They compare meta-agents, both unaided and equipped with an archive, against fixed methods built on synthetic practice and on corpus processing, across six heterogeneous benchmarks. The meta-agents decide for themselves how to prepare; the fixed methods apply one preparation strategy whatever the environment. A meta-agent variant achieves the highest Avg@3 reward on five of the six benchmarks. Fixed corpus processing remains best on the benchmark with the largest corpus.

The limits are part of the result. Larger study budgets do not reliably improve downstream reward, and the open-ended meta-agent does not win everywhere: where the corpus is largest, a fixed processing method still leads. The benefit that holds up is one of efficiency, in the sampling needed at test time rather than in the ceiling reached.

Computation, in this picture, is being moved rather than multiplied. Effort spent once, in a study phase before any task arrives, stands in for effort that would otherwise be spent again and again on repeated attempts at test time. A student who reads the textbook before the exam may not score higher than one who does not, but needs fewer tries to get there.