Gathering information about the environment state is the main goal in several planning tasks for autonomous agents, such as surveillance, inspection and tracking of objects. Such planning tasks are typically modeled using a Partially Observable Markov Decision Process (POMDP), and in the literature several approaches have emerged to consider information gathering during planning and execution. Similar developments can be seen in the field of active inference, which focuses on active information collection in order to be able to reach a goal. Both fields use POMDPs to model the environment, but the underlying principles for action selection are different. In this paper we create a bridge between both research fields by discussing how they relate to each other and how they can be used for information gathering. Our contribution is a tailored approach to model information gathering tasks directly in the active inference framework. A series of experiments demonstrates that our approach enables agents to gather information about the environment state. As a result, active inference becomes an alternative to common POMDP approaches for information gathering, which opens the door towards more cross cutting research at the intersection of both fields. This is advantageous, because recent advancements in POMDP solvers may be used to accelerate active inference, and the principled active inference framework may be used to model POMDP agents that operate in a neurobiologically plausible fashion.

Despite the significant advancements in computer vision models, their ability to generalize to novel object-attribute compositions remains limited. Existing methods for Compositional Zero-Shot Learning (CZSL) mainly focus on image classification. This paper aims to enhance CZSL in object detection without forgetting prior learned knowledge. We use Grounding DINO and incorporate Compositional Soft Prompting (CSP) into it and extend it with Compositional Anticipation. We achieve a 70.5% improvement over CSP on the harmonic mean (HM) between seen and unseen compositions on the CLEVR dataset. Furthermore, we introduce Contrastive Prompt Tuning to incrementally address model confusion between similar compositions. We demonstrate the effectiveness of this method and achieve an increase of 14.5% in HM across the pretrain, increment, and unseen sets. Collectively, these methods provide a framework for learning various compositions with limited data, as well as improving the performance of underperforming compositions when additional data becomes available.