Searcharxiv⌕ Search

arXiv subjects

Sehong Park

Publications and source records attributed to Sehong Park.

5 recordsLinked to original sources

The Race Is Decided Early! A Kernel Law Predicts Lock-In in Majority-Vote Test-Time RL

Test-time reinforcement learning with majority-vote rewards trains a language model on unlabeled questions by rewarding the samples that agree with its most frequent answer. During training, each question locks into one answer, on hard questions often a wrong one, but which answer it locks into is known only after training. In this paper, we show that this outcome is decided early, in a short race driven by sampling noise between the answers that are close at the start, and we derive the probability of each outcome. With a separate logit for each answer, the runner-up wins with a normal tail probability of the initial gap, scaled by an explicit noise constant. In a network, however, the answer logits depend on shared parameters, and through the kernel of these parameters the updates of the trailing answers also change the gap of the two leaders. This spillover shifts the boundary between the starts won by each leader, and a closed form computes the shifted probability from a kernel probe and the start logits. In LoRA experiments with three models from two families, predictions made before training with the kernel frozen at its start value forecast which answer each question locks into better than a logistic model refitted on the outcomes, although they use no fitted constants. The closed form matches these predictions on the runs in which a check before training finds that the two leaders stay above the other answers, and it explains why a law that ignores the spillover needs a strongly inflated step.

cs.LG↗

Endpoint capacity at strictly stable minimal boundaries:Gaussian layers and Schwarzschild deficits

Let $(M^n,\mathfrak g)$, $3\le n\le7$, be a one-ended asymptotically flat manifold whose boundary $Σ$ is a connected, strictly stable, outermost minimal hypersurface. As $p\downarrow1$, the $p$-capacity of $Σ$ converges to its area, by a theorem of Fogagnolo and Mazzieri. We establish a higher-order asymptotic expansion governed by a single function on $Σ$: the positive solution $g$ of the reciprocal Jacobi equation $Jg=g^{-1}$. After a universal term, the first correction is determined by the average of $\log g$, and the capacitary potential drops from one to zero across a Gaussian boundary layer of width $g\sqrt{p-1}$; under the normalized energy measure, the logarithm of the gradient has limiting variance $\frac12$ plus the variance of $\log g$. In dimension three, assuming in addition that $R_{\mathfrak g}+|\mathrm{II}|^2\ge0$ along $Σ$, we identify the capacity deficit of $Σ$ relative to the Schwarzschild horizon of the same area at order $p-1$: it is nonnegative, vanishes exactly when the induced metric of $Σ$ is round and $R_{\mathfrak g}+|\mathrm{II}|^2\equiv0$, and controls the deviation of $g$ from a constant through a relative-entropy identity.

math.DG↗

Charged Hawking Mass on CMC Collars with Charged Matter

We study strictly stable minimal horizons in three-dimensional time-symmetric charged initial data sets with cosmological constant, allowing charged matter. Along the constant-mean-curvature foliation issuing from the horizon through leaves with positive Jacobi operator, we prove exact first-variation identities for the charged Hawking mass. When the charge density has one sign compatible with the horizon charge, Hawking masses are nondecreasing, and equality between the two boundary values of the fixed-charge mass forces the entire collar to be isometric to an annulus in a source-free constant-curvature Reissner--Nordström--(anti-)de Sitter model. From this we recover the local rigidity results under a local maximality condition. At the minimal endpoint the defect entering these identities computes the fixed-charge Hawking-mass Hessian along the normalized lapse, and its vanishing characterizes the model horizon data.

gr-qc↗

Outward minimizing $p$-capacity, horizons, and Schwarzschild rigidity

Let $(M^3,g)$ be a complete Riemannian manifold diffeomorphic to $\R^3\setminus\{0\}$, with nonnegative scalar curvature. Assume that a distinguished end is asymptotically flat, with ADM mass $m_+$. For each $p\in(1,3)$, define $c_{O,p}$ as the infimum of the Schwarzschild-normalized $p$-capacity over outward-minimizing finite-perimeter boundaries separating the two ends. We prove that $m_+\ge c_{O,p}$ whenever $A(g)>0$, where $A(g)$ is the infimum of the areas of boundaries separating the two ends. Equality at a single exponent produces a least-area horizon, forces its exterior to be the Schwarzschild exterior of mass $m_+$, and yields equality at every exponent. In the equality case, if the second end is also asymptotically flat, its mass satisfies $m_-\ge m_+$, with equality precisely for the two-sided spatial Schwarzschild manifold. We also show that a strict gap between $c_{O,p}$ and the unconstrained capacity infimum $c_{M,p}$ detects a horizon.

math.DG↗