arXiv subjects

Xudong Wu

Publications and source records attributed to Xudong Wu.

11 recordsLinked to original sources

arXiv2026

EnergyBridge: Benchmarking Household Energy Management, User Participation, and Grid Flexibility

Residential virtual power plants (VPPs) can provide grid flexibility by shifting household demand, but physical flexibility becomes dependable capacity only when residents authorize a plan and the promised response is delivered. Existing benchmarks evaluate control but omit event-specific authorization. We present EnergyBridge, a benchmark and agent framework connecting capacity reporting, household authorization, and physical execution. It combines region-specific EnergyPlus environments for Tianjin and Berlin with an LLM-based User Participation Simulator. Against 584 persona- and event-matched human role-play judgments, the LLM-based User Participation Simulator preserves method ordering with a 5.3-point mean absolute acceptance error. Across conventional controllers and agent baselines, EnergyBridge achieves the highest simulated authorization, lowest event-window energy, and the most reliable capacity commitment in both regions. We release human data and codes for reproducible human-centered grid-flexibility research: https://github.com/Agentic-Intelligence-Lab/EnergyBridge.

cs.AI↗

arXiv2026

Distributionally Robust Listwise Preference Optimization

Existing robust preference optimization for language-model alignment mainly studies pairwise supervision and places robustness at the dataset, prompt, or preference-pair level. We instead study listwise preference optimization under ranking-label uncertainty: given a prompt and a candidate list, the observed ranking over that list may be ambiguous due to annotator inconsistency, near-ties, lossy rankwise feedback, or reward-model noise. We propose a pointwise total-variation robust Plackett--Luce objective that directly robustifies the ranking label conditional on the candidate list. The robust loss admits an exact decomposition into the nominal PL loss plus a worst-case PL correction, and the worst-case ranking is obtained by sorting current implicit scores in ascending order, reducing the inner maximization from $K!$ enumeration to $O(K\log K)$. This tractable structure yields strong offline and online optimization guarantees. In the offline fixed-list setting, the robust objective is convex and projected stochastic subgradient reaches global $\epsilon$-suboptimality with $O(\epsilon^{-2})$ sample complexity. In the online policy-induced setting, where candidate lists are generated by the current policy, we establish weak convexity and $\widetilde O(\epsilon^{-2})$ Moreau-envelope stationarity. Experiments in offline LLM alignment show that the proposed robust correction largely preserves performance under clean labels and improves robustness under noise. In online alignment, it makes reward-model-ranked candidate expansion more reliable and improves both reward-model and external GPT-4 judge metrics.

cs.AI↗

arXiv2026

On the Convergence of Self-Improving Online LLM Alignment

The Self-Improving Alignment (SAIL) algorithm addresses distribution shift by reducing a bilevel formulation of the problem to an efficient, single-level method. Empirically, SAIL has demonstrated strong performance on this task. However, a formal analysis of its convergence properties has been lacking. We identify a key theoretical challenge: the standard SAIL objective function is not guaranteed to be strongly concave due to unfavorable properties of its Hessian. To address this limitation, we propose a regularized objective, SAIL-RevKL, which incorporates a reverse Kullback-Leibler (KL) divergence penalty to improve the optimization landscape. Our central theoretical contribution is to prove that this regularized objective satisfies the Polyak-Lojasiewicz (PL) condition within a bounded parameter space. We establish global convergence guarantees, achieving a near-linear sample complexity. We further validate the effectiveness and stability of SAIL-RevKL through empirical evaluations, demonstrating that it outperforms the vanilla SAIL on both MuJoCo benchmarks and LLM alignment tasks.

cs.LG↗

arXiv2026

Core-Halo Decomposition: Decentralizing Large-Scale Fixed-Point Problems

We study solving large-scale fixed-point equation $x^\star=\bar F(x^\star)$ with decomposition. Standard strict decomposition assigns each agent a disjoint block and evaluates updates using only owned coordinates. For most operators, however, a block update may depend on variables outside the block. Truncating these dependencies by strict decomposition changes the mean operator and creates structural bias that cannot be removed by more samples, smaller stepsizes, or additional consensus. We therefore propose Core-Halo decomposition, which separates write ownership from read-only evaluation context: each agent updates its own core and reads from an overlapping halo. By aligning the Core-Halo decomposition with the block-dependence structure of $\bar F$, the original fixed-point problem can be implemented faithfully in a decentralized multi-agent system. We further characterize the fundamental obstruction faced by strict decomposition through a Bellman closure condition and a blockwise bias lower bound, showing that local-only updates can alter the original fixed-point operator. Finally, we conduct extensive experiments across a range of application settings, and demonstrate that Core-Halo achieves near-centralized performance while retaining the parallelism benefits of decentralization.

stat.ML↗

arXiv2026

A Lifting Theorem for Hybrid Classical-Quantum Communication Complexity

We investigates a model of hybrid classical-quantum communication complexity, in which two parties first exchange classical messages and subsequently communicate using quantum messages. We study the trade-off between the classical and quantum communication for composed functions of the form $f\circ G^n$, where $f:\{0,1\}^n\to\{\pm1\}$ and $G$ is an inner product function of $Θ(\log n)$ bits. To prove the trade-off, we establish a novel lifting theorem for hybrid communication complexity. This theorem unifies two previously separate lifting paradigms: the query-to-communication lifting framework for classical communication complexity and the approximate-degree-to-generalized-discrepancy lifting methods for quantum communication complexity. Our hybrid lifting theorem therefore offers a new framework for proving lower bounds in hybrid classical-quantum communication models. As a corollary, we show that any hybrid protocol communicating $c$ classical bits followed by $q$ qubits to compute $f\circ G^n$ must satisfy $c+q^2=Ω\big(\max\{\mathrm{deg}(f),\mathrm{bs}(f)\}\cdot\log n\big)$, where $\mathrm{deg}(f)$ is the degree of $f$ and $\mathrm{bs}(f)$ is the block sensitivity of $f$. For read-once formula $f$, this yields an almost tight trade-off: either they have to exchange $Θ\big(n\cdot\log n\big)$ classical bits or $\widetildeΘ\big(\sqrt n\cdot\log n\big)$ qubits, showing that classical pre-processing cannot significantly reduce the quantum communication required. To the best of our knowledge, this is the first non-trivial trade-off between classical and quantum communication in hybrid two-way communication complexity.

cs.CC↗

arXiv2026

Occupancy Reward Shaping: Improving Credit Assignment for Offline Goal-Conditioned Reinforcement Learning

The temporal lag between actions and their long-term consequences makes credit assignment a challenge when learning goal-directed behaviors from data. Generative world models capture the distribution of future states an agent may visit, indicating that they have captured temporal information. How can that temporal information be extracted to perform credit assignment? In this paper, we formalize how the temporal information stored in world models encodes the underlying geometry of the world. Leveraging optimal transport, we extract this geometry from a learned model of the occupancy measure into a reward function that captures goal-reaching information. Our resulting method, Occupancy Reward Shaping, largely mitigates the problem of credit assignment in sparse reward settings. ORS provably does not alter the optimal policy, yet empirically improves performance by 2.2x across 13 diverse long-horizon locomotion and manipulation tasks. Moreover, we demonstrate the effectiveness of ORS in the real world for controlling nuclear fusion on 3 Tokamak control tasks. Code: https://github.com/aravindvenu7/occupancy_reward_shaping; Website: https://aravindvenu7.github.io/website/ors/

cs.LG↗

arXiv2024

Almost Optimal Algorithms for Token Collision in Anonymous Networks

In distributed systems, situations often arise where some nodes each holds a collection of tokens, and all nodes collectively need to determine whether all tokens are distinct. For example, if each token represents a logged-in user, the problem corresponds to checking whether there are duplicate logins. Similarly, if each token represents a data object or a timestamp, the problem corresponds to checking whether there are conflicting operations in distributed databases. In distributed computing theory, unique identifiers generation is also related to this problem: each node generates one token, which is its identifier, then a verification phase is needed to ensure all identifiers are unique. In this paper, we formalize and initiate the study of token collision. In this problem, a collection of $k$ tokens, each represented by some length-$L$ bit string, are distributed to $n$ nodes of an anonymous CONGEST network in an arbitrary manner. The nodes need to determine whether there are tokens with an identical value. We present near optimal deterministic algorithms for the token collision problem with $\tilde{O}(D+k\cdot L/\log{n})$ round complexity, where $D$ denotes the network diameter. Besides high efficiency, the prior knowledge required by our algorithms is also limited. For completeness, we further present a near optimal randomized algorithm for token collision.

cs.DC↗

arXiv2022

Quantum Complexity of Weighted Diameter and Radius in CONGEST Networks

This paper studies the round complexity of computing the weighted diameter and radius of a graph in the quantum CONGEST model. We present a quantum algorithm that $(1+o(1))$-approximates the diameter and radius with round complexity $\widetilde O\left(\min\left\{n^{9/10}D^{3/10},n\right\}\right)$, where $D$ denotes the unweighted diameter. This exhibits the advantages of quantum communication over classical communication since computing a $(3/2-\varepsilon)$-approximation of the diameter and radius in a classical CONGEST network takes $\widetildeΩ(n)$ rounds, even if $D$ is constant [Abboud, Censor-Hillel, and Khoury, DISC '16]. We also prove a lower bound of $\widetildeΩ(n^{2/3})$ for $(3/2-\varepsilon)$-approximating the weighted diameter/radius in quantum CONGEST networks, even if $D=Θ(\log n)$. Thus, in quantum CONGEST networks, computing weighted diameter and weighted radius of graphs with small $D$ is strictly harder than unweighted ones due to Le Gall and Magniez's $\widetilde O\left(\sqrt{nD}\right)$-round algorithm for unweighted diameter/radius [PODC '18].

cs.DC↗

arXiv2022

Complexity of Eccentricities and All-Pairs Shortest Paths in the Quantum CONGEST Model

Computing the distance parameters of a network, including the diameter, radius, eccentricities and the all-pairs shortest paths (APSP) is a central problem in distributed computing. This paper investigates he dtistance parameters in the quantum CONGEST models and establishes almost linear lower bounds on eccentricities and APSP, which match the classical upper bounds. Our results imply that there is not quantum speedup for these two problems. In contrast with the diameter and radius, exchanging quantum messages is able to save the communication when the networks have low diameters [Le Gall and Magniez, PODC 2018]. We obtain the lower bounds via a reduction from the two-way quantum communication complexity of the set intersection [Razborov, Izvestiya Mathematics 2003].

quant-ph↗

arXiv2019

Evolving Influence Maximization in Evolving Networks

Influence Maximization (IM) aims to maximize the number of people that become aware of a product by finding the `best' set of `seed' users to initiate the product advertisement. Unlike prior arts on static social networks containing fixed number of users, we undertake the first study of IM in more realistic evolving networks with temporally growing topology. The task of evolving IM ({\bfseries EIM}), however, is far more challenging over static cases in the sense that seed selection should consider its impact on future users and the probabilities that users influence one another also evolve over time. We address the challenges through $\mathbb{EIM}$, a newly proposed bandit-based framework that alternates between seed nodes selection and knowledge (i.e., nodes' growing speed and evolving influences) learning during network evolution. Remarkably, $\mathbb{EIM}$ involves three novel components to handle the uncertainties brought by evolution:

cs.SI↗

arXiv2018

BLAG: Bandit On Large Action Set Graph

Information diffusion in social networks facilitates rapid and large-scale propagation of content. However, spontaneous diffusion behavior could also lead to the cascading of sensitive information, which is neglected in prior arts. In this paper, we present the first look into adaptive diffusion of sensitive information, which we aim to prevent from widely spreading without incurring much information loss. We undertake the investigation in networks with partially known topology, meaning that some users' ability of forwarding information is unknown. Formulating the problem into a bandit model, we propose BLAG (Bandit on Large Action set Graph), which adaptively diffuses sensitive information towards users with weak forwarding ability that is learnt from tentative transmissions and corresponding feedbacks. BLAG enjoys a low complexity of O(n), and is provably more efficient in the sense of half regret bound compared with prior learning method. Experiments on synthetic and three real datasets further demonstrate the superiority of BLAG in terms of adaptive diffusion of sensitive information over several baselines, with at least 40 percent less information loss, at least 10 times of learning efficiency given limited learning rounds and significantly postponed cascading of sensitive information.

cs.SI↗