arXiv subjects

Qiushui Xu

Publications and source records attributed to Qiushui Xu.

5 recordsLinked to original sources

arXiv2026

A Stochastic GDA Method With Backtracking For Solving Nonconvex Concave Minimax Problems

We propose a stochastic GDA (gradient descent ascent) method with backtracking (SGDA-B) to solve nonconvex-concave (NCC) minimax problems of the form: $\min_{\mathbf{x}} \max_y \sum_{i=1}^N g_i(x_i)+f(\mathbf{x},y)-h(y)$, where $h$ and $g_i$ for $i=1,\cdots,N$ are closed, convex functions, and for some $L,μ\geq 0$, $f$ is $L$-smooth and $f(\mathbf{x},\cdot)$ is $μ$-strongly concave for all $\mathbf{x}$ in the problem domain. We consider the stochastic setting where one only has an access to an unbiased stochastic oracle of $\nabla f$ with a finite variance bound $σ^2$. While most of the existing methods assume knowledge of $L$, $μ$ and/or $σ^2$, SGDA-B is agnostic to all of these problem parameters. Moreover, SGDA-B can support random block-coordinate updates. In the deterministic setting, i.e., $σ^2=0$ and one can compute $\nabla f$ exactly, SGDA-B can compute an $ε$-stationary point within $\mathcal{O}(Lκ^2/ε^2)$ and $\mathcal{O}(L^3/ε^4)$ gradient calls when $μ>0$ and $μ=0$, respectively, where $κ\triangleq L/μ$. In the stochastic setting, i.e., $σ^2>0$, for any $p\in(0,1)$ and $ε>0$, it can compute an $ε$-stationary point with high probability, which requires $\mathcal{O}(L κ^3 ε^{-4} \log^2(1/p))$ and $\tilde{\mathcal{O}}(L^4ε^{-7}\log^2(1/p))$ stochastic oracle calls, with probability at least $1-p$, when $μ>0$ and $μ=0$, respectively. To our knowledge, SGDA-B is the first GDA-type method with backtracking to solve NCC minimax problems and achieves the best complexity among the methods that are agnostic to $L$, $μ$ and $σ^2$. We also provide numerical results for SGDA-B on a distributionally robust learning problem illustrating the potential performance gains that can be achieved by SGDA-B.

math.OC↗

arXiv2026

In-Context Compositional Q-Learning for Offline Reinforcement Learning

Accurate estimation of the Q-function is a central challenge in offline reinforcement learning. However, existing approaches often rely on a shared global Q-function, which is inadequate for capturing the compositional structure of tasks that consist of diverse subtasks. We propose In-context Compositional Q-Learning (ICQL), an offline RL framework that formulates Q-learning as a contextual inference problem and uses linear Transformers to adaptively infer local Q-functions from retrieved transitions without explicit subtask labels. Theoretically, we show that, under two assumptions -- linear approximability of the local Q-function and accurate inference of weights from retrieved context -- ICQL achieves a bounded approximation error for the Q-function and enables near-optimal policy extraction. Empirically, ICQL substantially improves performance in offline settings, achieving gains of up to 16.4% on kitchen tasks and up to 8.8% and 6.3% on MuJoCo and Adroit tasks, respectively. These results highlight the underexplored potential of in-context learning for robust and compositional value estimation and establish ICQL as a principled and effective framework for offline RL.

cs.LG↗

arXiv2026

An Accelerated Primal Dual Algorithm with Backtracking for Decentralized Constrained Optimization

We propose a distributed accelerated primal-dual method with backtracking (D-APDB) for cooperative multi-agent constrained consensus optimization problems over an undirected network of agents, where only those agents connected by an edge can directly communicate to exchange large-volume data vectors using a high-speed, short-range communication protocol, e.g., WiFi, and we also assume that the network allows for one-hop simple information exchange beyond immediate neighbors as in LoRaWAN protocol. The objective is to minimize the sum of agent-specific composite convex functions over agent-specific private constraint sets. Unlike existing decentralized primal-dual methods that require knowledge of the Lipschitz constants, D-APDB automatically adapts to unknown smoothness constants by employing a distributed backtracking step-size search. Each agent relies only on first-order oracles associated with its own objective and constraint functions and on local communications with the neighboring agents, without any prior knowledge of Lipschitz constants. We establish $\mathcal{O}(1/K)$ convergence guarantees for sub-optimality, infeasibility and consensus violation, under standard assumptions on smoothness and on the connectivity of the communication graph. To our knowledge, when nodes have private constraints, especially when they are nonlinear convex constraints onto which projections are not cheap to compute, D-APDB is the first distributed method with backtracking that achieves the optimal convergence rate for the class of constrained composite convex optimization problems. We also provide numerical results for D-APDB on a distributed QCQP problem and distributed primal SVM training to illustrate the potential performance gains that can be achieved by D-APDB.

math.OC↗

arXiv2025

Unveiling Markov Heads in Pretrained Language Models for Offline Reinforcement Learning

Recently, incorporating knowledge from pretrained language models (PLMs) into decision transformers (DTs) has generated significant attention in offline reinforcement learning (RL). These PLMs perform well in RL tasks, raising an intriguing question: what kind of knowledge from PLMs has been transferred to RL to achieve such good results? This work first dives into this problem by analyzing each head quantitatively and points out Markov head, a crucial component that exists in the attention heads of PLMs. It leads to extreme attention on the last-input token and performs well only in short-term environments. Furthermore, we prove that this extreme attention cannot be changed by re-training embedding layer or fine-tuning. Inspired by our analysis, we propose a general method GPT2-DTMA, which equips a pretrained DT with Mixture of Attention (MoA), to accommodate diverse attention requirements during fine-tuning. Extensive experiments corroborate our theorems and demonstrate the effectiveness of GPT2-DTMA: it achieves comparable performance in short-term environments while significantly narrowing the performance gap in long-term environments.

cs.LG↗

arXiv2025

AGDA+: Proximal Alternating Gradient Descent Ascent Method with a Nonmonotone Adaptive Step-Size Search for Nonconvex Minimax Problems

We consider double-regularized nonconvex-strongly concave (NCSC) minimax problems of the form $(P):\min_{x\in\mathcal{X}} \max_{y\in\mathcal{Y}}g(x)+f(x,y)-h(y)$, where $g$, $h$ are closed convex, $f$ is $L$-smooth in $(x,y)$ and strongly concave in $y$. We propose a proximal alternating gradient descent ascent method AGDA+ that can adaptively choose nonmonotone primal-dual stepsizes to compute an approximate stationary point for $(P)$ without requiring the knowledge of the global Lipschitz constant $L$ and the concavity modulus $μ$. Using a nonmonotone step-size search (backtracking) scheme, AGDA+ stands out by its ability to exploit the local Lipschitz structure and eliminates the need for precise tuning of hyper-parameters. AGDA+ achieves the optimal iteration complexity of $\mathcal{O}(ε^{-2})$ and it is the first step-size search method for NCSC minimax problems that require only $3$ calls to $\nabla f$ on average per backtracking iteration. The numerical experiments demonstrate its robustness and efficiency.

math.OC↗