Motivation: Clustering is an unsupervised learning task with broad applications. Traditional clustering methods often rely on point estimates of model parameters, which can limit their ability to capture uncertainty. Bayesian clustering addresses this by incorporating uncertainty into parameter estimation. However, existing Bayesian inference methods like Markov Chain Monte Carlo and Variational Inference are computationally intensive and can produce biased approximations. To overcome these limitations, we propose Cluster-PFN, a transformer-based model inspired by Prior-Data Fitted Networks. Cluster-PFN simultaneously approximates the posterior distributions over the cluster assignments of individual data points and the total number of clusters for the given dataset in a single forward pass. Our model provides fast and accurate Bayesian clustering, supporting data with up to five features and conditioning on user-specified cluster counts.

Results: Our results demonstrate that Cluster-PFN can predict the number of clusters up to 20% more accurately than standard heuristics. It also outperforms the Bayesian Gaussian Mixture Model using Variational Inference (Bayesian GMM VI), achieving up to 60% higher scores on certain external metrics while being up to 20 times faster during inference. Additionally, Cluster-PFN surpasses both the traditional Gaussian Mixture Model and K-means++ across the same external evaluation metrics.

Large language models enable rich communication and flexible planning in embodied agents, yet their updates to internal state remain correlation driven, making them prone to hallucinations that undermine reliability in interactive settings. To address this limitation, we investigate whether a true causal model can replace template based belief, desire, and intention (BDI) updates and provide more stable reasoning. We introduce CausalMind, an embodied agent architecture that integrates a causal model trained on text based BDI transitions drawn from a synthetic dataset. We evaluate both the causal model and the full agent in iTHOR across solo and cooperative tasks.

The results indicate that learning causal variables from text embeddings is difficult. The model focused primarily on action related features, failing to separate belief, desire, and intention and instead overfitting to biases in the dataset. Expanding and refining the synthetic dataset improved performance but did not resolve the core issues, highlighting the need for richer data and more suitable encoders. At the agent level, perception errors from the vision module frequently led to hallucinated task completion, limiting the usefulness of teacher guidance and making complex tasks difficult in the iTHOR environment.

Although the current implementation does not outperform template based baselines, the study clarifies the obstacles that arise when extending causal representation learning to natural language BDI states. These findings outline the key requirements for future work toward embodied agents capable of robust causal reasoning rather than correlation driven updates.

Virtual agents have demonstrated remarkable progress in both competitive and cooperative en- vironments. Embodied agents, which enhance AI interactions with the physical world, show great promise for a variety of use cases in both virtual and non-virtual settings. This literature review ex- amines the intersection of embodied virtual agents with cognitive and social frameworks derived from human behavior, focusing on open-ended learning and intrinsic motivation. These paradigms, inspired by human learning and adaptability, offer a path to- wards addressing the limitations of current artificial systems. The literature survey provides a thorough analysis of the research landscape, discussing the definitions, applications, and benefits of embodied agents in virtual settings. Furthermore, it evaluates the methods and benchmarks used to assess the ca- pabilities of these agents, while offering possible solutions for developing the next generation of embodied agents.

This research paper aims to present how Theory of Mind (ToM) - the ability that allows humans to attribute mental states to others - can be used in the context of physically and virtually embodied computational agents. The focus is on using ToM for perspective-taking in environments with multiple computational agents interacting together. A Systematic Literature Review (SLR) was conducted providing a structured search process for collecting the literature. The findings from this review were obtained from 38 selected papers found in the literature since 2018. The review presented how computational ToM was inspired by the human ToM. Further, it summarised the current applications for perspective-taking for multi-agent settings, and it collected insights on high-level implementations of ToM agents. The findings highlight the complexity of ToM agents and the importance of ToM for agent interactions.

In the future, autonomous social robots are expected to seamlessly integrate into our society. To be perceived as interactive partners rather than mere tools, these robots must be embodied and capable of navigating complex, dynamic environments. This study explores the critical role of embodiment and examines the application of human-inspired socio-cognitive mechanisms in embodied agents. We conducted a systematic scoping review of 36 publications sourced from Scopus, IEEE Xplore, Web of Science, and PubMed. Our findings reveal a diverse array of human-inspired implementations in embodied agents, including curiosity-driven learning, categorical thinking, and predictive learning. Notably, human-like eye-gaze plays a crucial role in enhancing the anthropomorphism of these agents. By drawing inspiration from human interactions, we can transition from pre-programmed robots to fully autonomous agents that exhibit emergent behaviours and adapt to ever-changing conditions. Embodiment allows agents to communicate their intentions and desires through various modes, facilitating richer interactions. Additionally, multiple studies underscore the importance of using embodied agents to study human behaviours. Furthermore, the choice of embodiment type must depend on the task at hand; for some tasks, virtual embodiment is more effective, while for others, physical embodiment is preferable. This review provides a comprehensive starting point for future researchers interested in developing human-inspired embodied agents.

Continual learning (CL) enables intelligent systems to continually acquire, adapt, and apply knowledge, representing a dynamic paradigm in AI. For embodied agents—interacting with their environment physically and cognitively—CL enhances adaptability and reduces training costs significantly. In this literature review, we contribute by focusing on the application of CL in such agents, showcasing the approaches, means of evaluation and practical uses of this cognitive framework in real-world scenarios. We conclude that while CL holds promise for embodied agents, there exists a notable gap between the theoretical evaluation of CL and the complex real-world scenarios these agents operate in.

Active inference is a theory of the human brain characterising behaviour that minimises surprise. The free energy principle accounts for the adaptive behaviours of organisms through action, perception, and learning aimed at optimising reward or surprise. This study systematically reviews relevant literature to address their methodologies, relevance to mimicking social human behaviour, challenges, and limitations to guide future research by succinctly reporting previous findings and research gaps. Active inference models are extended with deep active inference, free-energy models, multimodal deep belief networks, predictive coding, and probabilistic programming. These models employ goal-directed, epistemic, reward-seeking, and decision-making behaviours and simulate cumulative culture. However, some of these models do not translate well to complex real-life applications due to their simplicity, computational demands, or the assumptions upon which they are based. Challenges with real-life applications include difficulty scaling to high-dimensional data and model simplicity. Furthermore, some experiments did not have enough data to validate or train their models.