# Reasoning by Superposition: A Theoretical Perspective on Chain of Continuous Thought

**Authors:** Hanlin Zhu, Shibo Hao, Zhiting Hu, Jiantao Jiao, Stuart Russell, Yuandong Tian  
**arXiv:** [2505.12514](https://arxiv.org/abs/2505.12514)  
**Date:** 2025-05-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.12514v1/

## Key points

- Large Language Models (LLMs) exhibit exceptional performance in various applications
- Chain-of-thoughts (CoTs) techniques within LLMs are utilized to tackle challenging reasoning problems
- Continuous CoTs outperform discrete CoTs in tasks like directed graph reachability
- A two-layer transformer with continuous CoTs can effectively solve the directed graph reachability problem
- Continuous thought vectors represent superposition states capable of encoding multiple search frontiers simultaneously, similar to parallel breadth-first search techniques
- Continuous CoTs naturally evolve during training to encode multiple search frontiers as superposition states without explicit guidance
- This research highlights the potential of continuous thought processes within LLMs for advancing reasoning capabilities across different domains

## Summary

In their paper titled "Reasoning by Superposition: A Theoretical Perspective on Chain of Continuous Thought," authors Hanlin Zhu, Shibo Hao, Zhiting Hu, Jiantao Jiao, Stuart Russell, and Yuandong Tian delve into the realm of Large Language Models (LLMs) and their exceptional performance in various applications. Specifically, they focus on the utilization of chain-of-thoughts (CoTs) techniques within LLMs to tackle challenging reasoning problems. These CoTs generate "thinking tokens" that aid in processing information before providing answers. While previous theoretical works have highlighted the effectiveness of CoTs with discrete tokens in enhancing LLM capabilities, recent advancements in continuous CoTs have shown superior performance without a comprehensive theoretical understanding. The authors particularly explore how continuous CoTs outperform their discrete counterparts in tasks such as directed graph reachability – a crucial graph reasoning problem with numerous practical applications. Through their research, the authors demonstrate that a two-layer transformer equipped with a specific number of steps of continuous CoTs can effectively solve the directed graph reachability problem. This breakthrough is significant as it surpasses the existing knowledge related to constant-depth transformers with discrete CoTs, which require significantly more decoding steps for similar outcomes. The key innovation lies in how each continuous thought vector represents a superposition state capable of encoding multiple search frontiers simultaneously. This approach mirrors parallel breadth-first search (BFS) techniques, enabling efficient exploration of various paths within the graph structure. In contrast, discrete CoTs are limited to selecting a single path from the superposition state, leading to sequential search processes that may result in prolonged solution times or local optima traps. Furthermore, extensive experiments conducted by the authors validate the theoretical framework proposed in their study. They observe that during training dynamics, continuous CoTs naturally evolve to encode multiple search frontiers as superposition states without explicit guidance. This emergent behavior enhances the model's ability to explore diverse paths concurrently and contributes to its overall reasoning prowess. Overall, this research sheds light on the intricate mechanisms underlying continuous thought processes within LLMs and highlights their potential for advancing reasoning capabilities across various domains. By bridging theoretical insights with empirical evidence, the authors pave the way for future developments in leveraging superposition states for enhanced problem-solving strategies within language models.