This work addresses a core limitation of current \textit{LLMs} when confronted with theoretical problems that cannot be directly inferred from information already present in their training data. It presents an overview of a method currently under investigation that aims to enhance the reasoning capabilities of \textit{LLMs} by breaking free from built-in patterns that constrain their ability to move beyond surface-level inference. When evaluating various \textit{LLMs}, particularly on graph-related problems, a consistent pattern emerged: one that reflects not only correctness issues but also limitations in reasoning strategy. The proposed method seeks to gradually shift the model’s behavior, encouraging divergence from repetitive response paths rather than supplying targeted prompts to elicit a single correct answer. Although a genuine breakthrough remains technically out of reach, the approach offers a promising direction toward overcoming current boundaries.