CST-Bench: Benchmarking Counterfactual Spatial Transition Reasoning with Causal Transition On-Policy Self-Distillation

Multimodal large language models (MLLMs) perform strongly on isolated spatial tasks, but whether their predictions remain persistent and mutually coherent across viewpoints and competing scenes is unclear. We formalize this behavioral property as spatial-state integrity and introduce CVT-Bench, a factorial diagnostic suite spanning two domains (CVT-Synthetic, CVT-Real), four context regimes (CVT-Isolated, irrelevant filler, attribute filler, CVT-Competing), three representations (Image, Text/BBox, Scene Graph), and ten viewpoint conditions (nine azimuths from $0^\circ$ to $360^\circ$ at $45^\circ$ increments, plus Top) with target views withheld. Across 200 scenes and 11,833 relational queries, we measure counterfactual accuracy, cycle consistency, the normalized survival coefficient, spatial realizability, and context-specific interference. Five state-of-the-art MLLMs exhibit rapid persistence loss and frequently produce jointly unrealizable states despite high local accuracy, with failures amplified by competing contexts and natural-scene complexity. Matched irrelevant and attribute filler controls show that prompt length, position, or loss of scene access alone are insufficient explanations. On average, Text/BBox improves accuracy, persistence, and realizability, while Scene Graph provides complementary gains; neither eliminates the instability. Thus, isolated spatial accuracy substantially overestimates robustness, establishing spatial-state integrity as a distinct evaluation target. The complete benchmark and codebase will be publicly released.

Existing spatial reasoning benchmarks mainly test spatial perception: reading off relations already visible in the input. Yet real-world spatial intelligence demands predictive spatial reasoning: constructing a scene from observations, anticipating how an intervention changes it, and reasoning about the unseen outcome. We introduce SpaceCast-Bench, the first benchmark to directly and diagnostically evaluate this capability. Built around an observe-transform-infer framework, its 3,862 questions from 182 real-world scenes span 16 task types at three levels: static perception, local prediction, and global prediction, progressively requiring scene understanding, spatial state updating, and relational inference over unobserved outcomes. Evaluating 21 models exposes a stark gap: the strongest model reaches only 58.0% against 87.2% human performance, while spatially specialized models remain near random chance. Controlled analyses further reveal that bridge views are critical for integrating distributed observations, and that explicit 3D evidence benefits models more reliably than generated outcome images or videos. Fine-tuning on our programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7% with macro-average gains across six out-of-domain benchmarks.

Although Multimodal Large Language Models (MLLMs) have made substantial progress, their spatial reasoning may still produce intermediate judgments inconsistent with the input image, allowing errors to propagate through the reasoning chain and affect the final answer. Existing methods mainly improve spatial reasoning through training or additional spatial information, without considering whether the reasoning process itself is faithful to the model input. Our study shows that unfaithful reasoning chains significantly reduce final-answer accuracy. To address this issue, we propose a modular and training-free framework for spatial reasoning verification and correction. The framework constructs a Spatial Evidence Graph (SEG), which associates atomic spatial evidence extracted from Chain-of-Thought reasoning with visual entities, spatial relations, source steps, and visual evidence. Spatial Evidence Reliability Assessment (SERA) evaluates the reliability of visual evidence based on object existence, localization, and geometric measurements. The framework then identifies the earliest spatial evidence unit contradicted by reliable visual evidence and guides the original MLLM to revise the subsequent reasoning and final answer. Across 15 model-dataset settings, our method achieves an average accuracy of 68.94%, outperforming the compared baselines by 8.55 percentage points on average. Our code will be open-sourced.