arXiv subjects

Zhuzhong Qian

Publications and source records attributed to Zhuzhong Qian.

7 recordsLinked to original sources

arXiv2026

Failure-Guided Co-Evolution of Prompts and Training Data

Automatic prompt optimization (APO) improves language-model programs by revising prompts from task feedback, yet it typically holds its training data fixed. Repeatedly optimizing against the same instances confines feedback to weaknesses already represented in those data, leaving related failure conditions unexplored. We therefore view each failure as a dual signal: it indicates both how the prompt should be revised and what new training evidence should be synthesized. We introduce FORGE, a failure-guided framework that co-evolves prompts and training data. FORGE abstracts imperfect executions into reusable failure modes and synthesizes new training data through four complementary mutation strategies. Verified instances are fed back into prompt search, allowing updated prompts to expose the next data needs. Across eight heterogeneous benchmarks, FORGE improves the aggregate score over the unoptimized baseline by 16.52 percentage points and outperforms all evaluated APO baselines. The synthesized data also transfer beyond FORGE: in a transfer study, they improve all nine APO comparisons by 2--9 points and all three GRPO comparisons by 4--8 points under matched optimization budgets. These results establish failures as a shared interface between prompt optimization and data synthesis, and show the benefit of jointly adapting what a model is instructed to do and what it learns from.

cs.SE↗

arXiv2026

Denoise to Track: Harnessing Video Diffusion Priors for Robust Correspondence

In this work, we introduce HeFT (Head-Frequency Tracker), a zero-shot point tracking framework that leverages the visual priors of pretrained video diffusion models. To better understand how they encode spatiotemporal information, we analyze the internal representations of Video Diffusion Transformer (VDiT). Our analysis reveals that attention heads act as minimal functional units with distinct specializations for matching, semantic understanding, and positional encoding. Additionally, we find that the low-frequency components in VDiT features are crucial for establishing correspondences, whereas the high-frequency components tend to introduce noise. Building on these insights, we propose a head- and frequency-aware feature selection strategy that jointly selects the most informative attention head and low-frequency components to enhance tracking performance. Specifically, our method extracts discriminative features through single-step denoising, applies feature selection, and employs soft-argmax localization with forward-backward consistency checks for correspondence estimation. Extensive experiments on TAP-Vid benchmarks demonstrate that HeFT achieves state-of-the-art zero-shot tracking performance, approaching the accuracy of supervised methods while eliminating the need for annotated training data. Our work further underscores the promise of video diffusion models as powerful foundation models for a wide range of downstream tasks, paving the way toward unified visual foundation models.

cs.CV↗

arXiv2026

From Off-Policy to On-Policy: Enhancing GUI Agents via Bi-level Expert-to-Policy Assimilation

Vision-language models are increasingly deployed as computer-use agents (CUAs) that operate desktops and browsers. Top-performing CUAs are framework-based systems that decompose planning and execution, while end-to-end screenshot-to-action policies are easier to deploy but lag behind on benchmarks such as OSWorld-Verified. GUI datasets like OSWorld pose two bottlenecks: they expose only a few hundred interactive, verifiable tasks and environments, and expert trajectories must be gathered by interacting with these environments, making such data hard to scale. We therefore ask how reinforcement learning from verifiable rewards (RLVR) can best exploit a small pool of exist expert trajectories to train end-to-end policies. Naively mixing these off-policy traces into on-policy RLVR is brittle: even after format conversion, expert trajectories exhibit structural mismatch and distribution shift from the learner. We propose BEPA (Bi-Level Expert-to-Policy Assimilation), which turns static expert traces into policy-aligned guidance via self-rolled reachable trajectories under the base policy (LEVEL-1) and a per-task, dynamically updated cache used in RLVR (LEVEL-2). On OSWorld-Verified, BEPA improves UITARS1.5-7B success from 22.87% to 32.13% and raises a held-out split from 5.74% to 10.30%, with consistent gains on MMBench-GUI and Online-Mind2Web. Our code and data are available at: https://github.com/LEON-gittech/Verl_GUI.git

cs.AI↗

arXiv2025

Optimizing Cross-Client Domain Coverage for Federated Instruction Tuning of Large Language Models

Federated domain-specific instruction tuning (FedDIT) for large language models (LLMs) aims to enhance performance in specialized domains using distributed private and limited data, yet identifying key performance drivers and optimal augmentation strategies remains challenging. We empirically establish that cross-client domain coverage, rather than data heterogeneity, is the pivotal factor. We then introduce FedDCA, an algorithm that explicitly maximizes this coverage through diversity-oriented client center selection and retrieval-based augmentation, constructing diverse, non-redundant cross-client instruction sets. Extensive experiments across multiple domains demonstrate FedDCA's superiority over eleven baselines, achieving performance gains of up to 29.19\% and domain coverage improvements of 4.82\%-21.36\%. FedDCA maintains its effectiveness in diverse and challenging scenarios, including data selection, held-out settings where task-specific public data is scarce and various data heterogeneity, with manageable privacy risks. This work clarifies critical FedDIT dynamics and presents FedDCA as an effective, privacy-preserving, and scalable solution for advancing domain-specific LLM tuning.

cs.LG↗

arXiv2018

PingAn: An Insurance Scheme for Job Acceleration in Geo-distributed Big Data Analytics System

Geo-distributed data analysis in a cloud-edge system is emerging as a daily demand. Out of saving time in wide area data transfer, some tasks are dispersed to the edges. However, due to limited computing, overload interference and cluster-level unreachable troubles, efficient execution in the edges is hard, which obstructs the guarantee on the efficiency and reliability of jobs. Launching copies across clusters can be an insurance on a task's completion. Considering cluster heterogeneity and accompanying remote data fetch, cluster selection of copies affects execution quality, as different insuring plans drive different revenues. For providing On-Line-Real-Time analysis results, a system needs to insure the geo-distributed resource for the arriving jobs. Our challenge is to achieve the optimal revenue by dynamically weighing the gains due to insurance against the loss of occupying extra resource for insuring. To this end, we design PingAn, an online insurance algorithm promising $(1+\varepsilon)\!\!-\!speed \ o(\frac{1}{\varepsilon^2+\varepsilon})\!\!-\!competitive$ in sum of the job flowtimes via cross-cluster copying for tasks. PingAn shares resource among the anterior fraction of jobs with the least unprocessed datasize and the fraction is adjustable to fit the system load condition. After sharing, PingAn concretely insures for tasks following efficiency-first reliability-aware principle to optimize the revenue of copies on jobs' performance. Trace-driven simulations demonstrate that PingAn can reduce the average job flowtimes by at least $14\%$ than the state-of-the-art speculation mechanisms. We also build PingAn in Spark on Yarn System to verify its practicality and generality. Experiments show that PingAn can reduce the average job flowtimes by up to $40\%$ comparing to the default Spark execution.

cs.DC↗

arXiv2018

Towards Reliable (and Efficient) Job Executions in a Practical Geo-distributed Data Analytics System

Geo-distributed data analytics are increasingly common to derive useful information in large organisations. Naive extension of existing cluster-scale data analytics systems to the scale of geo-distributed data centers faces unique challenges including WAN bandwidth limits, regulatory constraints, changeable/unreliable runtime environment, and monetary costs. Our goal in this work is to develop a practical geo-distribued data analytics system that (1) employs an intelligent mechanism for jobs to efficiently utilize (adjust to) the resources (changeable environment) across data centers; (2) guarantees the reliability of jobs due to the possible failures; and (3) is generic and flexible enough to run a wide range of data analytics jobs without requiring any changes. To this end, we present a new, general geo-distributed data analytics system, HOUTU, that is composed of multiple autonomous systems, each operating in a sovereign data center. HOUTU maintains a job manager (JM) for a geo-distributed job in each data center, so that these replicated JMs could individually and cooperatively manage resources and assign tasks. Our experiments on the prototype of HOUTU running across four Alibaba Cloud regions show that HOUTU provides nearly efficient job performance as in the existing centralized architecture, and guarantees reliable job executions when facing failures.

cs.DC↗

arXiv2018

RDMAvisor: Toward Deploying Scalable and Simple RDMA as a Service in Datacenters

RDMA is increasingly adopted by cloud computing platforms to provide low CPU overhead, low latency, high throughput network services. On the other hand, however, it is still challenging for developers to realize fast deployment of RDMA-aware applications in the datacenter, since the performance is highly related to many lowlevel details of RDMA operations. To address this problem, we present a simple and scalable RDMA as Service (RaaS) to mitigate the impact of RDMA operational details. RaaS provides careful message buffer management to improve CPU/memory utilization and improve the scalability of RDMA operations. These optimized designs lead to simple and flexible programming model for common and knowledgeable users. We have implemented a prototype of RaaS, named RDMAvisor, and evaluated its performance on a cluster with a large number of connections. Our experiment results demonstrate that RDMAvisor achieves high throughput for thousand of connections and maintains low CPU and memory overhead through adaptive RDMA transport selection.

cs.DC↗