Adaptive user interfaces must infer each user’s latent goals from implicit behavioural cues a capability closely related to Theory of Mind (ToM), the capacity to attribute beliefs, desires, and intentions to others from observable actions. Existing simulation benchmarks for adaptive user interfaces either depend on pre-collected human interaction data or lack support for systematic evaluation under diverse, previously unseen user behaviours; at the same time, existing interface agents treat user intent as implicit, limiting robustness when behaviour deviates from training conditions. This thesis introduces AssistUI, a JAX-based cooperative simulation benchmark that formalises user–assistant interaction as a hidden-goal problem, drawing on assistance games, ad hoc teamwork (AHT), and multi-agent reinforcement learning (MARL). In a clothing-selection task with asymmetric information, an interface agent must deduce the user’s hidden target outfit solely from click-and-reject behaviour, without access to the user’s goal. Within this benchmark, we augment a standard Independent Proximal Policy Optimisation (IPPO) agent with a recurrent goal-inference head a GRU-based module that maintains an explicit belief distribution over the user’s goal and conditions the policy on this belief. The intention-aware agent is evaluated against diverse synthetic user populations generated by Fictitious Co-Play (FCP), Best-Response Diversity (BRDiv), and Conventions via Mixed-Play Diversity (CoMeDi), as well as hand-coded slot-constrained user policies that model behavioural biases never seen during training. Across complementary experiment groups spanning specialisation, distribution-shift robustness, crossplay coordination, in-distribution performance, and population diversity, we find that a well-trained recurrent agent can achieve competitive cross-play performance without explicit goal inference or pipeline retraining, that goal inference does not uniformly improve generalisation and can be detrimental for well-trained baselines, that the relationship between measured population diversity and cross-play robustness is weak, and that existing population-generation methods can produce functionally compatible partners even when abstract diversity metrics remain low. These results suggest that intention-aware architectures and training-population design are complementary rather than interchangeable tools for building robust adaptive user interfaces.