Chain of Thought (CoT) prompting has been shown to significantly improve the performance of large language models (LLMs), particularly in arithmetic and reasoning tasks, by instructing the model to produce intermediate reasoning steps. Despite the remarkable empirical success of CoT and its theoretical advantages in enhancing expressivity, the mechanisms underlying CoT training remain largely unexplored. In this paper, we study the training dynamics of transformers over a CoT objective on an in-context weight prediction task for linear regression. We prove that while a one-layer linear transformer without CoT can only implement a single step of gradient descent (GD) and fails to recover the ground-truth weight vector, a transformer with CoT prompting can learn to perform multi-step GD autoregressively, achieving near-exact recovery. Furthermore, we show that the trained transformer effectively generalizes on the unseen data. With our technique, we also show that looped transformers significantly improve final performance compared to transformers without looping in the in-context learning of linear regression. Empirically, we demonstrate that CoT prompting yields substantial performance improvements.

Armijo line-search (Armijo-LS) is a standard method to set the step-size for gradient descent (GD). For smooth functions, Armijo-LS alleviates the need to know the global smoothness constant $L$ and adapts to the local smoothness, enabling GD to converge faster. However, existing theoretical analyses of GD with Armijo-LS (GD-LS) do not characterize this fast convergence. We show that if the objective function satisfies a certain non-uniform smoothness condition, GD-LS converges provably faster than GD with a constant $1/L$ step-size (denoted as GD(1/L)). Our results imply that for convex losses corresponding to logistic regression and multi-class classification, GD-LS can converge to the optimum at a linear rate and, hence, improve over the sublinear convergence of GD(1/L). Furthermore, for non-convex losses satisfying gradient domination (for example, those corresponding to the softmax policy gradient in RL or generalized linear models with a logistic link function), GD-LS can match the fast convergence of algorithms tailored for these specific settings. Finally, we prove that under the interpolation assumption, for convex losses, stochastic GD with a stochastic line-search can match the fast convergence of GD-LS.

The inductive bias and generalization properties of large machine learning models are -- to a substantial extent -- a byproduct of the optimization algorithm used for training. Among others, the scale of the random initialization, the learning rate, and early stopping all have crucial impact on the quality of the model learnt by stochastic gradient descent or related algorithms. In order to understand these phenomena, we study the training dynamics of large two-layer neural networks. We use a well-established technique from non-equilibrium statistical physics (dynamical mean field theory) to obtain an asymptotic high-dimensional characterization of this dynamics. This characterization applies to a Gaussian approximation of the hidden neurons non-linearity, and empirically captures well the behavior of actual neural network models. Our analysis uncovers several interesting new phenomena in the training dynamics: $(i)$ The emergence of a slow time scale associated with the growth in Gaussian/Rademacher complexity; $(ii)$ As a consequence, algorithmic inductive bias towards small complexity, but only if the initialization has small enough complexity; $(iii)$ A separation of time scales between feature learning and overfitting; $(iv)$ A non-monotone behavior of the test error and, correspondingly, a `feature unlearning' phase at large times.

Gradient-based methods are well-suited for derivative-free optimization (DFO), where finite-difference (FD) estimates are commonly used as gradient surrogates. Traditional stochastic approximation methods, such as Kiefer-Wolfowitz (KW) and simultaneous perturbation stochastic approximation (SPSA), typically utilize only two samples per iteration, resulting in imprecise gradient estimates and necessitating diminishing step sizes for convergence. In this paper, we first explore an efficient FD estimate, referred to as correlation-induced FD estimate, which is a batch-based estimate. Then, we propose an adaptive sampling strategy that dynamically determines the batch size at each iteration. By combining these two components, we develop an algorithm designed to enhance DFO in terms of both gradient estimation efficiency and sample efficiency. Furthermore, we establish the consistency of our proposed algorithm and demonstrate that, despite using a batch of samples per iteration, it achieves the same convergence rate as the KW and SPSA methods. Additionally, we propose a novel stochastic line search technique to adaptively tune the step size in practice. Finally, comprehensive numerical experiments confirm the superior empirical performance of the proposed algorithm.

Differentially Private Stochastic Gradient Descent (DPSGD) is widely used to protect sensitive data during the training of machine learning models, but its privacy guarantees often come at the cost of model performance, largely due to the inherent challenge of accurately quantifying privacy loss. While recent efforts have strengthened privacy guarantees by focusing solely on the final output and bounded domain cases, they still impose restrictive assumptions, such as convexity and other parameter limitations, and often lack a thorough analysis of utility. In this paper, we provide rigorous privacy and utility characterization for DPSGD for smooth loss functions in both bounded and unbounded domains. We track the privacy loss over multiple iterations by exploiting the noisy smooth-reduction property and establish the utility analysis by leveraging the projection's non-expansiveness and clipped SGD properties. In particular, we show that for DPSGD with a bounded domain, (i) the privacy loss can still converge without the convexity assumption, and (ii) a smaller bounded diameter can improve both privacy and utility simultaneously under certain conditions. Numerical results validate our results.

This paper provides a theoretical analysis of a new learning problem for recommender systems where users provide feedback by comparing pairs of items instead of rating them individually. We assume that comparisons stem from latent user and item features, which reduces the task of predicting preferences to learning these features from comparison data. Similar to the classical matrix factorization problem, the main challenge in this learning task is that the resulting loss function is nonconvex. Our analysis shows that the loss function exhibits (restricted) strong convexity near the true solution, which ensures gradient-based methods converge exponentially, given an appropriate warm start. Importantly, this result holds in a sparse data regime, where each user compares only a few pairs of items. Our main technical contribution is to extend certain concentration inequalities commonly used in matrix completion to our model. Our work demonstrates that learning personalized recommendations from comparison data is computationally and statistically efficient.

Deep neural networks trained using gradient descent with a fixed learning rate $\eta$ often operate in the regime of "edge of stability" (EOS), where the largest eigenvalue of the Hessian equilibrates about the stability threshold $2/\eta$. In this work, we present a fine-grained analysis of the learning dynamics of (deep) linear networks (DLNs) within the deep matrix factorization loss beyond EOS. For DLNs, loss oscillations beyond EOS follow a period-doubling route to chaos. We theoretically analyze the regime of the 2-period orbit and show that the loss oscillations occur within a small subspace, with the dimension of the subspace precisely characterized by the learning rate. The crux of our analysis lies in showing that the symmetry-induced conservation law for gradient flow, defined as the balancing gap among the singular values across layers, breaks at EOS and decays monotonically to zero. Overall, our results contribute to explaining two key phenomena in deep networks: (i) shallow models and simple tasks do not always exhibit EOS; and (ii) oscillations occur within top features. We present experiments to support our theory, along with examples demonstrating how these phenomena occur in nonlinear networks and how they differ from those which have benign landscape such as in DLNs.

In this paper, we apply quantum machine learning (QML) to predict the stock prices of multiple assets using a contextual quantum neural network. Our approach captures recent trends to predict future stock price distributions, moving beyond traditional models that focus on entire historical data, enhancing adaptability and precision. Utilizing the principles of quantum superposition, we introduce a new training technique called the quantum batch gradient update (QBGU), which accelerates the standard stochastic gradient descent (SGD) in quantum applications and improves convergence. Consequently, we propose a quantum multi-task learning (QMTL) architecture, specifically, the share-and-specify ansatz, that integrates task-specific operators controlled by quantum labels, enabling the simultaneous and efficient training of multiple assets on the same quantum circuit as well as enabling efficient portfolio representation with logarithmic overhead in the number of qubits. This architecture represents the first of its kind in quantum finance, offering superior predictive power and computational efficiency for multi-asset stock price forecasting. Through extensive experimentation on S\&P 500 data for Apple, Google, Microsoft, and Amazon stocks, we demonstrate that our approach not only outperforms quantum single-task learning (QSTL) models but also effectively captures inter-asset correlations, leading to enhanced prediction accuracy. Our findings highlight the transformative potential of QML in financial applications, paving the way for more advanced, resource-efficient quantum algorithms in stock price prediction and other complex financial modeling tasks.

Figure 1: In a challenging real-world experiment with limited communication (agents can only exchange information every 30 seconds), our method RAMEN enables each turtlebot to successfully map the full scene while only physically visiting half of the scene (explored areas and trajectories are colored accordingly). Our method achieves accuracy comparable to the ground truth while the baseline method (DiNNO) fails to converge. Abstract --Multi-agent neural implicit mapping allows robots to collaboratively capture and reconstruct complex environments with high fidelity. However, existing approaches often rely on synchronous communication, which is impractical in real-world scenarios with limited bandwidth and potential communication interruptions. This paper introduces RAMEN: Real-time Asynchronous Multi-agEnt Neural implicit mapping, a novel approach designed to address this challenge. RAMEN employs an uncertainty-weighted multi-agent consensus optimization algorithm that accounts for communication disruptions. When communication is lost between a pair of agents, each agent retains only an outdated copy of its neighbor's map, with the uncertainty of this copy increasing over time since the last communication. Using gradient update information, we quantify the uncertainty associated with each parameter of the neural network map. Neural network maps from different agents are brought to consensus on the basis of their levels of uncertainty, with consensus biased towards network parameters with lower uncertainty. T o achieve this, we derive a weighted variant of the decentralized consensus alternating direction method of multipliers (C-ADMM) algorithm, facilitating robust collaboration among agents with varying communication and update frequencies.

We study robust parameter-efficient fine-tuning (PEFT) techniques designed to improve accuracy and generalization while operating within strict computational and memory hardware constraints, specifically focusing on large-language models (LLMs). Existing PEFT methods often lack robustness and fail to generalize effectively across diverse tasks, leading to suboptimal performance in real-world scenarios. To address this, we present a new highly computationally efficient framework called AdaZo-SAM, combining Adam and Sharpness-Aware Minimization (SAM) while requiring only a single-gradient computation in every iteration. This is achieved using a stochastic zeroth-order estimation to find SAM's ascent perturbation. We provide a convergence guarantee for AdaZo-SAM and show that it improves the generalization ability of state-of-the-art PEFT methods. Additionally, we design a low-rank gradient optimization method named LORENZA, which is a memory-efficient version of AdaZo-SAM. LORENZA utilizes a randomized SVD scheme to efficiently compute the subspace projection matrix and apply optimization steps onto the selected subspace. This technique enables full-parameter fine-tuning with adaptive low-rank gradient updates, achieving the same reduced memory consumption as gradient-low-rank-projection methods. We provide a convergence analysis of LORENZA and demonstrate its merits for pre-training and fine-tuning LLMs.