Spatial reasoning is a fundamental ability of multimodal large language models (MLLMs), yet most existing benchmarks focus on static scene understanding. When objects are moved or swapped, a model must infer both spatial relations and how the intervention alters the scene state. To address this gap we introduce CST-Bench, a benchmark dedicated to counterfactual spatial reasoning. It contains 7,082 examples covering six intervention types and twelve transformation operators, and provides dense causal annotations such as intervention sources, affected relations, expected effects, invariant relations, and metric changes. We evaluate general‑purpose MLLMs on CST-Bench and compare post‑training strategies on this benchmark and on standard spatial benchmarks. Results show that methods successful on ordinary spatial tasks can still fail under counterfactual interventions. We formulate counterfactual spatial reasoning as causal state‑transition modeling and propose CT‑OPSD, which leverages CST‑Bench annotations to guide a teacher model in learning counterfactual state transitions and causal reasoning chains, then distills these chains into a student model that reasons only from real observations. Experiments establish CST‑Bench as a challenging and reusable supervision resource for counterfactual spatial reasoning, and CT‑OPSD provides an initial demonstration of how its annotations can improve models.\n\nReview