Zero-Shot Chain-of-Thought Reasoning Guided by Evolutionary Algorithms in Large Language Models

Feihu Jin, Yifan Liu, Ying Tan

Introduction

Large Language Models (LLMs) have demonstrated remarkable reasoning capabilities, particularly when prompted with instructions like "Let’s think step by step." This technique, known as Chain-of-Thought (CoT) prompting, encourages models to decompose complex problems into intermediate steps. However, standard zero-shot CoT typically relies on a static, "one-size-fits-all" prompt. This approach ignores the fact that LLMs are highly sensitive to prompt phrasing; a prompt optimized for arithmetic may be suboptimal for symbolic logic or commonsense reasoning.

To address this limitation, researchers have introduced Evolution-of-Thought (EoT). This method treats prompt generation as an optimization problem solved through an Evolutionary Algorithm (EA), where the LLM itself acts as the optimizer. By dynamically generating and selecting the best prompt for every individual problem instance, EoT significantly enhances model performance without requiring human-labeled examples.

The Evolution-of-Thought (EoT) Framework

The core of the EoT approach is a four-stage evolutionary process that allows the LLM to refine its own reasoning instructions:

  1. Initialization: The process begins with a small population of basic CoT prompts, such as standard step-by-step instructions and more detailed planning-oriented prompts.
  2. Evolution (Crossover & Mutation): The LLM is tasked with performing meta-cognitive operations. It "crossovers" existing prompts to create hybrids and "mutates" them to generate a diverse pool of new prompting strategies tailored to the problem domain.
  3. Selection: For a specific question $Q$, the LLM evaluates the generated pool and selects the prompt it deems most suitable for that particular instance.
  4. Problem Rewriting & Execution: Using the selected prompt, the LLM first rewrites the original question to ensure a full understanding of variables and constraints. It then executes the reasoning steps to produce the final answer.
Comparison of standard Zero-shot CoT versus the proposed Zero-shot EoT framework.

This instance-aware approach ensures that the model's reasoning path is specifically calibrated for the complexity and type of the question at hand.

Experimental Evaluation

The researchers evaluated EoT across ten diverse reasoning benchmarks, categorized into three domains:

  • Arithmetic: MultiArith, GSM8K, AddSub, AQuA, SingleEq, and SVAMP.
  • Commonsense: CommonsenseQA (CSQA) and StrategyQA.
  • Symbolic Reasoning: Last Letter Concatenation and Coin Flip.

The primary experiments utilized GPT-3.5-turbo, with additional testing on GPT-4 to verify the method's scalability to more advanced models.

Results and Analysis

EoT achieved an average accuracy of 83.5% across mathematical tasks, outperforming standard Zero-shot CoT (80.7%) and even surpassing Few-shot Manual-CoT (82.2%), which relies on eight human-crafted examples. This suggests that automated, instance-specific evolution can be more effective than static human expertise.

Key Performance Highlights:

  • Mathematical Reasoning: On the AQuA dataset, accuracy improved from 55.3% to 62.2%. On SVAMP, EoT pushed GPT-4 performance from 88.7% to 92.7%.
  • Symbolic Reasoning: As shown in the table below, EoT consistently outperformed other zero-shot baselines and rivaled or exceeded few-shot performance on tasks like Last Letter Concatenation.
Performance comparison on symbolic reasoning tasks.

Insights into Model Behavior

The study revealed several critical insights. First, LLMs are capable "prompt engineers," successfully performing crossover and mutation tasks. Second, the "rewriting" step is vital; by restating the problem under the guidance of the evolved prompt, the model minimizes hallucinations and calculation errors. Finally, the researchers observed a positive correlation between population size ($N$) and accuracy, indicating that a larger pool of mutated prompts increases the likelihood of finding an optimal reasoning path.

Conclusion

Evolution-of-Thought (EoT) represents a significant step forward in zero-shot prompting. By leveraging the internal optimization capabilities of LLMs, it moves away from static instructions toward dynamic, instance-aware reasoning. While the method incurs higher computational costs due to multiple LLM calls per question, its ability to outperform human-crafted few-shot prompts demonstrates the immense potential of self-evolving AI systems. Future research may explore optimizing these prompts in continuous embedding spaces or applying the framework to smaller, open-source models.

Summary updated

BibTeX
@inproceedings{jin2025zero,
  author = {Feihu Jin and Yifan Liu and Ying Tan},
  title = {{Zero-Shot Chain-of-Thought Reasoning Guided by Evolutionary Algorithms in Large Language Models}},
  booktitle = {Nlpcc},
  year = {2025},
  doi = {10.1007/978-981-95-3352-7\_5},
}
Figure