Reproduce “Aligned Textual Scoring Rules” 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
Yuxuan Lu, Yifan Wu, Jason Hartline, Michael J. Curry
An Aligned Scoring Rule (ASR) for textual information elicitation is introduced, achieving provable truthfulness and strong alignment with human preferences. Evaluated on peer grading datasets, ASR significantly reduces Mean Squared Error by over 50% compared to non-aligned truthful baselines, while also demonstrating high Pearson (0.717) and Spearman (0.622) correlations with instructor scores.
Key insights
- It is possible to design textual scoring rules that simultaneously possess provable truthfulness (properness) and strong alignment with human preferences.
- Large Language Models can serve as effective and scalable proxies for human judgment, with LLM-Judge scores showing high correlation (0.5540 Pearson) with human instructor scores.
- Optimizing within the convex space of separate scoring rules allows for efficient and interpretable learning of proper scoring rules that minimize misalignment with reference scores.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.