When Emotional Stimuli meet Prompt Designing: An Auto-Prompt Graphical Paradigm

Chenggian Ma, Xiangyu Zhao, Chunhui Zhang, Yanzhao Qin, Wentao ZhangarXiv 2024 · DOI · arXiv

Introduction

Prompt engineering has evolved into two primary branches: Framework Prompts, which utilize structured reasoning such as Chain-of-Thought or Tree-of-Thoughts, and Stimulating Prompts, which leverage emotional or psychological cues to nudge models toward better performance. Despite their individual successes, current approaches often suffer from high manual design effort and a lack of systematic integration between structured logic and emotional stimuli.

Comparison of prompt engineering paradigms across key dimensions.

To bridge this gap, we introduce the Auto-Prompt Graphical Paradigm (APGP). APGP automates the creation of hybrid prompts, providing a "prompt-free" user experience while executing complex, emotionally-charged reasoning under the hood. By combining graph-based reasoning with systematic emotional layering, the system significantly improves accuracy on complex logical tasks.

Methodology: The Stimuli Graphical Processor

The core of APGP is the Stimuli Graphical Processor, which conceptualizes problem-solving as a structured graph containing both immutable (fixed) and variable components.

The Auto-Prompt Graphical Paradigm (APGP) workflow, illustrating the transition from problem definition to final answer through iterative validation.

The framework follows a multi-step process:

  1. Problem Abstraction: The model first defines the problem precisely, a technique inspired by "Take a Step Back" prompting, to avoid jumping to premature conclusions.
  2. Diversified Generation: Three distinct solutions are generated in parallel to ensure fault tolerance and diversity of thought.
  3. Thought Aggregation: Instead of simple voting, the LLM performs "biological hybridization," synthesizing the strengths of all three solutions into a single optimal version.
  4. Self-Verification & Backtracking: The model validates its output. If a failure is detected, it extracts "experience" from the error and backtracks to an earlier node in the reasoning graph to generate a new solution.
  5. Emotional Layering: Throughout these steps, fixed system prompts incorporate emotional stimuli—such as capitalization ("YOU ARE THE CLEVEREST!"), exclamation marks, and encouraging phrases—to maximize the model's latent capabilities.

Experimental Evaluation

The APGP framework was evaluated on two rigorous benchmarks:

  • Ruozhiba: A Chinese dataset consisting of logical traps, puns, and brain teasers. APGP achieved a 62.08% success rate, demonstrating high linguistic adaptability and reasoning depth.
  • BIG-Bench Hard (BBH): A suite of 23 challenging tasks focusing on multi-step reasoning and logic. Ablation studies confirmed that the inclusion of emotional stimuli consistently improved performance across diverse domains, from formal fallacies to multi-step arithmetic.

Key Insights

  • The "Encouragement" Factor: LLMs, trained on massive human corpora, exhibit human-like responses to psychological praise and pressure. Systematic use of encouraging tones leads to measurable performance gains.
  • Aggregation over Voting: Merging multiple "thoughts" into a hybrid solution proves more effective than traditional majority-rule voting, as it allows the model to combine complementary strengths.
  • Graph-Based Robustness: Treating reasoning as a graph allows for efficient error recovery through targeted backtracking rather than full restarts, enhancing the system's overall robustness.

Conclusion and Future Work

APGP demonstrates that the systematic combination of structured logic and emotional stimuli significantly boosts LLM performance on complex reasoning tasks. Future research will focus on addressing the "Judge" problem—where models may confidently validate incorrect answers—and optimizing the computational cost of traversing complex reasoning graphs for simpler queries.

Summary updated

BibTeX
@misc{ma2024when,
  author = {Chenggian Ma and Xiangyu Zhao and Chunhui Zhang and Yanzhao Qin and Wentao Zhang},
  title = {{When Emotional Stimuli meet Prompt Designing: An Auto-Prompt Graphical Paradigm}},
  year = {2024},
  eprint = {2404.10500},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.10500},
}
Figure