Proper Scoring Rules for Text and Super-alignment
summary
Proposes an algorithm that converts instructor reviews into a numeric rubric by calling summarization and question-answering oracles (domain-knowledge-free queries to ChatGPT), maps peer and instructor reviews to binary vectors, then applies an optimized multidimensional scoring rule — typically a learned V-shaped scoring rule — to maximize incentive for effort. Claims provable approximate properness (error scales linearly with QA mistakes) and adversarial robustness to prompt injections, tests on peer‑grading with ~10 instructor-reviewed examples and learns weights where “algorithm correctness” dominates, improving over quadratic scoring.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:30Motivation for scoring rules
- 5:49Proper scoring rules for text
- 12:00Numerical scoring for peer grading
- 20:04Optimizing scoring for effort
- 29:27From numerical to textual scoring
- 34:33Theoretical guarantees with oracles
- 38:58Aligning scoring with human evaluation
- 46:04Empirical results and future work
- 48:11Q&A
related talks
33mThe Silent Crash:Why Your RAG Evaluation Metrics Are Lying to You
46mEliciting and Enforcing Subjective Individual Fairness
34mBuilding Trust with LLMs
42mAdvancements in Evaluating Large Language Model Applications
55mThe ABCs of Scoring: Why, How, and Getting a Good Grade
46mBeyond BLEU & ROUGE — Evaluating LLMs and AI Systems
15mVulnerability of Text-Matching in ML/AI Conference Reviewer Assignments...
29mAnnotating data the right way