To CoT or To Loop? A Formal Comparison Between Chain-of-Thought and Looped Transformers
This work offers practical cues for choosing between reasoning paradigms in AI, addressing a foundational problem for researchers in machine learning and reasoning systems.
The paper tackled the comparative capabilities of Chain-of-Thought (CoT) and Looped Transformers by providing a formal analysis, showing that Looped Transformers efficiently simulate parallel computations for deterministic tasks, while CoT excels at approximate inference for compositional structures.
Chain-of-Thought (CoT) and Looped Transformers have been shown to empirically improve performance on reasoning tasks and to theoretically enhance expressivity by recursively increasing the number of computational steps. However, their comparative capabilities are still not well understood. In this paper, we provide a formal analysis of their respective strengths and limitations. We show that Looped Transformers can efficiently simulate parallel computations for deterministic tasks, which we formalize as evaluation over directed acyclic graphs. In contrast, CoT with stochastic decoding excels at approximate inference for compositional structures, namely self-reducible problems. These separations suggest the tasks for which depth-driven recursion is more suitable, thereby offering practical cues for choosing between reasoning paradigms.