Wednesday 19 March 2025
Artificial intelligence has made tremendous strides in recent years, but despite its many accomplishments, language models still struggle to accurately reason and solve complex problems. A team of researchers has proposed a novel approach that leverages human psychology to improve the performance of these large language models (LLMs).
The idea is simple: by incorporating elements of human reasoning into LLMs, they can be trained to better identify mistakes and correct them on their own. The resulting system, dubbed PSSD, is designed to mimic the way humans think and reason when faced with complex problems.
At its core, PSSD relies on a multi-agent debate paradigm that involves three distinct roles: intuition-based id, rule-driven superego, and script-centric ego. Each role serves a specific purpose, such as providing initial attempts or summarizing rules, which helps to refine the LLM’s thinking process.
To test PSSD, researchers used two datasets: AdvHotpotQA, which contains complex questions that require reasoning and analysis, and 2WikiMultiHopQA, which includes questions that span multiple articles. They compared the performance of PSSD against a baseline system called Self-Contrast, which also uses multi-agent debate but lacks the human psychology-inspired approach.
The results were impressive: PSSD outperformed Self-Contrast in terms of accuracy and stability, correctly answering 96 percent of questions compared to 89 percent for Self-Contrast. Moreover, PSSD was able to accurately identify missing answers and provide more correct predictions than incorrect ones.
But what’s truly remarkable about PSSD is its ability to learn from mistakes. By incorporating human psychology into the system, it can recognize when it’s made an error and refine its thinking process accordingly. This leads to a significant reduction in incorrect answers and a marked improvement in overall performance.
The implications of this research are far-reaching. As LLMs become increasingly sophisticated, they will be tasked with solving more complex problems that require nuanced understanding and reasoning. PSSD offers a promising approach for improving the accuracy and reliability of these systems, which could have significant benefits in fields such as healthcare, finance, and education.
Of course, there are still many challenges to overcome before PSSD can be widely adopted. For one, it requires a significant amount of training data, which can be difficult to obtain, especially for complex domains like medicine or law.
Cite this article: “Artificial Intelligence Enhanced with Human Psychology for Improved Reasoning and Problem-Solving”, The Science Archive, 2025.
Artificial Intelligence, Language Models, Human Psychology, Reasoning, Complex Problems, Multi-Agent Debate, Intuition, Rule-Driven, Script-Centric, Error Reduction
