Item Response Scaling Laws: A Measurement Theory Approach for Efficient and Generalizable Neural Scaling Estimation

In three linesIRSL integrates Item Response Theory (IRT) into LM scaling laws, reducing complexity from O(M×N) to O(M+N). Validated on 6,612 checkpoints and 37,682 questions, Beta-IRT achieves comparable accuracy with 99.9% fewer questions (50 per benchmark).

Summary generated by Claude — human-verified