Scholar - SciOpen

Spatial reasoning, defined as the ability to infer and compose relations among entities—is fundamental to geoinformatics applications such as spatial querying and map understanding. However, existing works such as Bidirectional Encoder Representations from Transformers (BERT)-based spatial Question Answering (QA) models and neuro-symbolic models rely on dataset-specific patterns, leading to shortcut learning, where reliance on superficial lexical cues rather than true relational understanding. Recent Large Language Models (LLMs)-based works, including fine-tuning and Chain-of-Thought (CoT) prompting, partially alleviate shortcut learning but remain limited by non-causal reasoning, where predictions depend on spurious correlations rather than stable relational structure. To address these limitations, we propose $C a u s a l I n f e r e n c e a n d R e a s o n i n g v i a C o m p a c t s U b n e t w o r k I d e n T i f i c a t i o n (C I R C U I T - X)$ , motivated by the hypothesis that spatial reasoning in LLMs is governed by compact causal parameter subsets (a.k.a causal circuits). $C I R C U I T - X$ operates in two stages: (i) Causal Importance Estimation (Stage I) via structured interventions to mitigate shortcut learning, and (ii) Minimal Circuit Discovery (Stage II) via structured pruning to mitigate non-causal reasoning. Empirically, $C I R C U I T - X$ achieves up to 91% accuracy on SPAtial Reasoning on Textual Question Answering (SPARTQA) and 87% on StepGame, outperforming State-of-the-Art (SOTA) methods while improving intervention robustness by up to +11% and causal consistency by up to +16%. Therefore, it retains up to 96% of full-model performance using only 3%–6% of active parameters, while demonstrating strong robustness under cross-domain transfer with improvements of up to +10% in accuracy and substantially higher intervention stability under distribution shifts.