Hwang et al.
It turns self-consistency training into a quantitative prediction: which answer a question will lock into can be computed before training, from the starting logits and a kernel probe, with no fitted constants. For readers interested in the probability of test-time reasoning it is a clean example of a diffusion-style race analysed in closed form.
- Prerequisites
- Gaussian tail probabilities, diffusion approximations of noisy gradient updates, and the neural tangent kernel.
- Focus on
- The two-answer race with separate logits first, then the kernel spillover between answers and its closed form; the LoRA experiments last.
- Reading time
- About 2 hours
arXiv:2611.11165
Zhilova
Normal approximation over convex sets is what justifies the high-dimensional bootstrap and simultaneous confidence regions. The paper gives explicit constants and shows that four moments are the threshold: bounds on the fourth coordinate moments give a worst-case error linear in \(d\), and higher moments do not improve the rate.
- Prerequisites
- Lindeberg and Stein-type normal approximation, Gaussian smoothing of indicator functions, and Rademacher sums.
- Focus on
- The main theorems and the lower-bound construction first, then the Gaussian-smoothing argument for Rademacher sums on Parseval frames and the random swaps that extend it to general sums.
- Reading time
- About 3 hours
arXiv:2611.13106