Abstract
This paper evaluates the scientific reasoning capabilities of GPT-4o using the "Google-proof" GPQA dataset, testing seven different prompt engineering techniques. The researchers discovered a surprising "accuracy-explanation paradox": while techniques like Self-Consistency yield the highest accuracy (52.99%), they often produce the poorest logical explanations, suggesting that LLMs frequently rely on pattern recognition rather than true logical inference.
Introduction
While Large Language Models (LLMs) have become remarkably adept at general conversation and summarization, their ability to perform "hard" scientific reasoning—the kind required in medicine, law, or high-level physics—remains a "black box." Most benchmarks focus solely on whether a model gets the right answer (accuracy). However, in critical fields, why a model reached a conclusion is just as important as the conclusion itself. This study addresses the gap between performance and interpretability, investigating whether sophisticated prompting actually improves a model's logical "thinking" or just its ability to guess correctly.
Methodology
The study utilized GPT-4o and applied seven distinct prompting strategies: Direct Answer (Zero-Shot), Chain-of-Thought (CoT), Zero-Shot CoT, Self-Ask, Self-Consistency, Decomposition, and a custom Multipath Prompting method. The evaluation was performed on the GPQA (Graduate-Level Google-Proof Q&A) dataset, specifically the "main" set of 448 questions written by experts in biology, physics, and chemistry. To measure the quality of reasoning, the authors used Cosine Similarity to compare the model's generated explanations against ground-truth explanations provided by human experts.
Key Insights and Results
The Accuracy-Explanation Paradox
A primary finding is the clear trade-off between accuracy and explanation quality. Simple prompts (Direct Answer, CoT) resulted in explanations that most closely matched human expert logic, even if they weren't always the most accurate. Conversely, Self-Consistency achieved the highest accuracy at 52.99% but performed the second-worst in explanation similarity. This suggests that LLMs often "hallucinate" a reasoning path to fit a pattern they recognize, rather than performing first-principles logical deduction.
Subdomain Sensitivity and Performance Stability
GPT-4o's performance is not uniform across scientific fields. It performed significantly better in Organic Chemistry (reaching up to 80% accuracy) than in Physics or Genetics. Despite these variations, the model's performance was found to be highly stable.
As shown in the analysis of accumulated correctness, the model's strengths and weaknesses in specific scientific subdomains are consistent. The researchers found that GPT-4o's performance was highly stable, with an R-squared value greater than 0.99 when looking at the linear accumulation of correct answers over the question set.
Key Performance Metrics
- Top Performer (Accuracy): Self-Consistency (52.99%).
- Worst Performer (Accuracy): Standard Chain-of-Thought (43.75%), suggesting that few-shot examples might bias the model in high-level contexts.
- Reasoning Quality: Direct Answer and Zero-Shot CoT provided the most "human-like" scientific reasoning paths.
Conclusion and Future Work
The findings indicate that while prompting can boost accuracy, it does not necessarily improve the underlying logical reasoning. Future work will focus on developing hybrid AI approaches that integrate structured reasoning frameworks and involving domain experts to provide qualitative grades on the model's logic, moving beyond automated similarity scores.
Summary updated