Reproduce “Cognitive Biases in LLM-Assisted Software Development” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinyi Zhou, Zeinadsadat Saghi, Sadra Sabouri, Rahul Pandita, Mollie McGuire, Souti Chattopadhyay
A study from the University of Southern California and GitHub Research systematically identifies and categorizes cognitive biases manifesting in Large Language Model (LLM)-assisted software development, revealing that LLM interactions significantly increase the likelihood of biased actions and subsequent rework. The research introduces a new taxonomy of 15 specific bias categories and proposes mitigation strategies for developers and LLM tool designers.
Key insights
- LLM-related actions are significantly more likely to be associated with cognitive biases (53.71%) and have higher reversal rates (29.46%) compared to overall development actions (47.14% biased, 25% reversed).
- Traditional cognitive bias categories were found to be insufficient to fully explain the patterns of biased decision-making in LLM-assisted contexts.
- Biases like 'Instant Gratification' (prioritizing speed over quality) and 'Suggester Preference' (over-trusting AI suggestions) are particularly prevalent in LLM interactions, leading to frequent rework.
- Developers exhibit awareness of these biases, with over 50% perceiving 14 out of 15 categories to have negative impacts on their work.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.