notesum.ai
Published at November 26Inference Scaling $\scriptsize\mathtt{F}$Laws: The Limits of LLM Resampling with Imperfect Verifiers
cs.LG
cs.AI
Released Date: November 26, 2024
Authors: Benedikt Stroebl1, Sayash Kapoor1, Arvind Narayanan1
Aff.: 1Princeton University

| Parameter | Value |
| Probability of correct solution (easy task), | 0.87 |
| Probability of correct solution (hard task), | 0.13 |
| Completeness, | 1 |
| Soundness, | 0.75 |
| Prior probability of easy task, | 0.58 |
| Prior probability of hard task, | 0.42 |
| Benefit for correct solution (true positive), | 1 |
| Cost for false positive, | [-0.7, -1, -2] |
| Computational cost per attempt, | 0 |