We introduce BabyTheorist, a benchmark designed to evaluate whether a machine can acquire reusable theories solely from observation. The key ability is observational theory learning: extracting underlying compositional programs from pure before‑after visual observations and applying them to novel situations.\n\nBabyTheorist continuously streams pairs of visual states generated by hidden compositional programs. During training, learners only see the resulting transitions, without any program labels, task identifiers, or support sets. To facilitate theory learning and evaluation, the benchmark controls the program language, a curriculum that gradually increases program complexity, recurring frequent sub‑programs, and a continual train‑test protocol with level skipping.\n\nSeven representative baselines plateau well below the ceiling, and the performance gap widens as the curriculum advances, indicating that current program‑learning approaches have not yet addressed this observational theory learning capability.\n\nReview: BabyTheorist offers a rigorous and scalable environment for studying machine observation‑based reasoning, highlighting the shortcomings of existing models in abstract theory construction.