Study Cram Scheduling and Learning Assistant using Mixture-of-Recursions (MoR) in Agentic AI

The digital learning industry has become a rapidly growing sector since early 2020. Traditionally, before the expansive presence of Artificial Intelligence (AI), browsing has been the primary source of information for learning or attending quizzes on educational websites. However, recently, Large Language Models (LLMs) have been utilized, which has led to efficiency competition among multiple AI models. LLMs have handled multiple workloads, which has reduced time efficiency. Agentic AI has introduced a divide-and-conquer strategy to improve time efficiency. This approach has supported the learning domain by enabling tasks to be completed effectively while reducing overall computation. Advancements in Transformer architecture is presented at the token level computation as well as in other areas resulting in the development and have been introduced as Mixture-of-Recursions (MoR). This approach has aimed to dynamically change the number of routing operations based on the importance assigned to tokens. MoR also represents an advancement of existing recursion mechanisms. The MoR architecture has been compared with Vanilla and other recursion architectures to evaluate improvements in accuracy. Considering the pretrained checkpoints of MoR, Vanilla, and Recursion as foundation, fine-tuning has been conducted, with each architecture incorporating a quiz model, a scheduler model, a pedagogy model, and an orchestrator model. The orchestrator models of MoR, Vanilla, and Recursion have been evaluated using metrics such as routing confidence margin (RCM), decision stability, token efficiency, and router entropy. This advancement has examined the possibility of maintaining higher accuracy when deployed as an orchestrator for effectively conveying commands to agents on an academic platform. The RCM of the MoR model has an almost 40-fold increase in RCM over Vanilla, demonstrating agents are being routed with greater decisiveness and reliability.