Inference-Aware Meta-Alignment of LLMs via Non-Linear GRPO

Aligning large language models (LLMs) to diverse human preferences is fundamentally challenging since criteria can often conflict with each other. Inference-time alignment methods have recently gained popularity as they allow LLMs to be aligned to multiple criteria via different alignment algorithms at inference time. However, inference-time alignment is computationally expensive since it often requires multiple forward passes of the base model. In this work, we propose inference-aware meta-alignment (IAMA), a novel approach that enables LLMs to be aligned to multiple criteria with limited computational budget at inference time. IAMA trains a base model such that it can be effectively aligned to multiple tasks via different inference-time alignment algorithms. To solve the non-linear optimization problems involved in IAMA, we propose non-linear GRPO, which provably converges to the optimal solution in the space of probability measures.

A Relative-Budget Theory for Reinforcement Learning with Verifiable Rewards in Large Language Model Reasoning

Reinforcement learning (RL) is a dominant paradigm for improving the reasoning abilities of large language models, yet its effectiveness varies across tasks and compute budgets. We propose a \emph{relative-budget} theory explaining this variation through a single quantity called relative budget $ξ:= H/\mathbb{E}[T]$, where $H$ is the generation horizon (token budget) and $T$ denotes the number of tokens until the first correct solution under a base policy. We show that $ξ$ determines sample efficiency by controlling reward variance and the likelihood of informative trajectories. Our analysis reveals three regimes: in the \emph{deficient} regime ($ξ\to 0$), informative trajectories are rare and the sample complexity explodes; in the \emph{balanced} regime ($ξ=Θ(1)$), informative trajectories occur with non-negligible probability and RL is maximally sample-efficient; and in the \emph{ample} regime ($ξ\to \infty$), learning remains stable but marginal gains per iteration diminish. We further provide finite-sample guarantees for online RL that characterize learning progress across these regimes. Specifically, in a case study under idealized distributional assumptions, we show that the relative budget grows linearly over iterations. Our empirical results confirm these predictions in realistic settings, identifying a budget $ξ\in [1.5, 2.0]$ that maximizes learning efficiency and coincides with peak reasoning performance.

Optimal Variance and Covariance Estimation under Differential Privacy in the Add-Remove Model and Beyond

In this paper, we study the problem of estimating the variance and covariance of datasets under differential privacy in the add-remove model. While estimation in the swap model has been extensively studied in the literature, the add-remove model remains less explored and more challenging, as the dataset size must also be kept private. To address this issue, we develop efficient mechanisms for variance and covariance estimation based on the \emph{B\'{e}zier mechanism}, a novel moment-release framework that leverages Bernstein bases. We prove that our proposed mechanisms are minimax optimal in the high-privacy regime by establishing new minimax lower bounds. Moreover, beyond worst-case scenarios, we analyze instance-wise utility and show that the B\'{e}zier-based estimator consistently achieves better utility compared to alternative mechanisms. Finally, we demonstrate the effectiveness of the B\'{e}zier mechanism beyond variance and covariance estimation, showcasing its applicability to other statistical tasks.

FedDuA: Doubly Adaptive Federated Learning

Federated learning is a distributed learning framework where clients collaboratively train a global model without sharing their raw data. FedAvg is a popular algorithm for federated learning, but it often suffers from slow convergence due to the heterogeneity of local datasets and anisotropy in the parameter space. In this work, we formalize the central server optimization procedure through the lens of mirror descent and propose a novel framework, called FedDuA, which adaptively selects the global learning rate based on both inter-client and coordinate-wise heterogeneity in the local updates. We prove that our proposed doubly adaptive step-size rule is minimax optimal and provide a convergence analysis for convex objectives. Although the proposed method does not require additional communication or computational cost on clients, extensive numerical experiments show that our proposed framework outperforms baselines in various settings and is robust to the choice of hyperparameters.

Accelerating Differentially Private Federated Learning via Adaptive Extrapolation

The federated learning (FL) framework enables multiple clients to collaboratively train machine learning models without sharing their raw data, but it remains vulnerable to privacy attacks. One promising approach is to incorporate differential privacy (DP)-a formal notion of privacy-into the FL framework. DP-FedAvg is one of the most popular algorithms for DP-FL, but it is known to suffer from the slow convergence in the presence of heterogeneity among clients' data. Most of the existing methods to accelerate DP-FL require 1) additional hyperparameters or 2) additional computational cost for clients, which is not desirable since 1) hyperparameter tuning is computationally expensive and data-dependent choice of hyperparameters raises the risk of privacy leakage, and 2) clients are often resource-constrained. To address this issue, we propose DP-FedEXP, which adaptively selects the global step size based on the diversity of the local updates without requiring any additional hyperparameters or client computational cost. We show that DP-FedEXP provably accelerates the convergence of DP-FedAvg and it empirically outperforms existing methods tailored for DP-FL.

Mean-field Analysis on Two-layer Neural Networks from a Kernel Perspective

In this paper, we study the feature learning ability of two-layer neural networks in the mean-field regime through the lens of kernel methods. To focus on the dynamics of the kernel induced by the first layer, we utilize a two-timescale limit, where the second layer moves much faster than the first layer. In this limit, the learning problem is reduced to the minimization problem over the intrinsic kernel. Then, we show the global convergence of the mean-field Langevin dynamics and derive time and particle discretization error. We also demonstrate that two-layer neural networks can learn a union of multiple reproducing kernel Hilbert spaces more efficiently than any kernel methods, and neural networks acquire data-dependent kernel which aligns with the target function. In addition, we develop a label noise procedure, which converges to the global optimum and show that the degrees of freedom appears as an implicit regularization.

Approximation and Estimation Ability of Transformers for Sequence-to-Sequence Functions with Infinite Dimensional Input

Despite the great success of Transformer networks in various applications such as natural language processing and computer vision, their theoretical aspects are not well understood. In this paper, we study the approximation and estimation ability of Transformers as sequence-to-sequence functions with infinite dimensional inputs. Although inputs and outputs are both infinite dimensional, we show that when the target function has anisotropic smoothness, Transformers can avoid the curse of dimensionality due to their feature extraction ability and parameter sharing property. In addition, we show that even if the smoothness changes depending on each input, Transformers can estimate the importance of features for each input and extract important features dynamically. Then, we proved that Transformers achieve similar convergence rate as in the case of the fixed smoothness. Our theoretical results support the practical success of Transformers for high dimensional data.