Training smarter AI agents does not require more teacher output — just better-placed output.
Researchers studying how to fine-tune large language model agents found that the standard approach, behavioral cloning from full teacher demonstrations, creates a mismatch: the model trains on contexts the teacher reached, not the contexts the student will actually encounter. Recent fixes have tried filtering those teacher outputs for quality. This paper argues that filtering is the wrong lever. Instead, the researchers framed training data construction as a budget problem: given fixed compute, where should teacher inference actually go? Testing across three benchmarks — HotpotQA, ALFWorld, and Terminal-Bench-Dev — they found that short, unfiltered teacher continuations placed at student-reached contexts beat both full demonstrations and filtered alternatives at matched cost.
The finding matters because inference costs real money, and every step a teacher model generates burns it. If a few short continuations at the right moments match or exceed elaborate filtering pipelines, the practical implication is that expensive curation workflows may be solving the wrong problem entirely. The result also implies that where supervision is placed in a trajectory may matter more than how much of it there is.
The research adds a data point to a growing debate about whether quality filtering in synthetic training data is worth its overhead — a debate with stakes beyond academic benchmarks, given how heavily frontier labs now rely on distillation from stronger models to train cheaper ones.