Reproduce “Surface Form Competition: Why the Highest Probability Answer Isn't Always Right” 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
Ari Holtzman, Peter West, Vered Shwartz, Yejin Choi, Luke Zettlemoyer
A new scoring function, Domain Conditional Pointwise Mutual Information (PMI_DC), improves zero-shot multiple-choice performance of large language models by addressing "surface form competition," where conceptual answers distribute probability across various textual forms. The approach consistently outperforms standard methods on 16 datasets across GPT-2 and GPT-3 models.
Key insights
- Raw probability assigned by generative large language models (LLMs) to a string does not directly translate to conceptual correctness in a constrained selection task due to "surface form competition".
- Factoring out the unconditional likelihood of a string, as done by Domain Conditional Pointwise Mutual Information (PMI_DC), is crucial for accurately assessing conceptual plausibility in multiple-choice scenarios.
- The "COPA Flipped" experiment provides mechanistic evidence that surface form competition is the primary reason for performance degradation in standard LLM scoring methods for multiple-choice tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.