Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 10·5 min read
In-Context Examples Suppress Scientific Knowledge Recall in LLMs
TL;DR
In-context examples can hinder large language models' (LLMs) ability to recall scientific knowledge, shifting their focus from knowledge-driven reasoning to empirical pattern fitting. This effect was observed across 60 tasks in five scientific domains, indicating a cautionary note for practitioners using LLMs in scientific applications.
✦ Why It Matters
Evaluate the use of in-context examples in LLMs to ensure they support rather than undermine scientific reasoning.
Key Takeaways
- Test LLMs without in-context examples for scientific tasks.
- Monitor accuracy changes when using in-context examples.
- Consider the balance between empirical fitting and knowledge recall.
- Adjust training strategies based on model performance in specific domains.
How It Works
The study demonstrates that when in-context examples are provided, LLMs tend to prioritize empirical pattern fitting over their pretrained scientific knowledge. This shift occurs even when the examples are derived from the same scientific formulas, leading to a reliance on observed data rather than theoretical understanding.