Abstract Understanding whether artificial agents can perform intuitive reasoning that is similar to that observed in human infants, is a key challenge in AI. The Baby Intuitions Benchmark (BIB) was designed to test such capabilities through tasks that evaluate an agent’s ability to recognize physical expectations. While the IRENE model has achieved strong results on BIB using graph-based spatial representations, this thesis investigates an alternative approach: replacing graphs with Spatial Semantic Pointers (SSP), a biologically inspired, high-dimensional vector representation rooted in the Semantic Pointer Architecture (SPA). We propose a novel model architecture that combines SSPs with sequential encoders such as Transformers and Long-Short-Term-Memory (LSTM). Two input strategies are explored —trajectory-based and path-integrated encodings— to assess whether SSPs generalize across diverse scenarios. Performance is evaluated using the Violation of Expectation (VoE) paradigm, which measures a model’s ability to detect unexpected behaviour. Our results show, that Transformer-based models outperform LSTM counterparts, and that trajectory-based input yields better performance than path-integrated representations. Although the models did not surpass state-of-the-art performance on any individual BIB task, they demonstrate that SSPs can serve as a viable alternative to graph-based inputs.