arXiv subjects

Yanwei Xu

Publications and source records attributed to Yanwei Xu.

16 recordsLinked to original sources

arXiv2026

TuneAhead: Predicting Fine-tuning Performance Before Full Training Begins

Fine-tuning large language models (LLMs) is compute-intensive and error-prone: model performance depends sensitively on data quality and hyperparameter choices, and na\"ive runs can even degrade model performance. This raises a practical question:can we predict fine-tuning performance before committing to a full training run? We present TUNEAHEAD, a lightweight framework for pre-hoc prediction of fine-tuning performance. TUNEAHEAD encodes each candidate run as a meta-feature vector that combines static dataset descriptors with dynamic probe features from a short standardized probe. A predictor maps these features to performance estimates, while SHAP-based attributions provide interpretable diagnostics that reveal which specific features drive the prediction. Across 1,300+ fine-tuning runs on Qwen2.5-7B-Instruct, TUNEAHEAD consistently outperforms strong baselines such as Early-Stop Extrapolation and ProxyLM. On a held-out test set of 370 runs, TUNEAHEAD achieves an RMSE of 1.47 percentage points and places 95.1% of predictions within +3/-3 percentage points of the true score. These accurate continuous predictions support practical go/no-go screening policies that can reduce unnecessary full fine-tuning while retaining most promising runs.

cs.LG↗

arXiv2026

AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle

Scientific research has traditionally been human-intensive, requiring researchers to coordinate literature, ideas, experiments, manuscripts, and review responses across long project cycles. The rise of LLM-based scientific agents creates an opportunity to automate this process. Such a system must support the full research lifecycle, maintain structured persistent memory across projects, and improve its own research procedures over time. However, existing systems either partially satisfy or fail to satisfy these requirements, leaving a gap for a unified automated scientific research system. As a result, we present AutoSci, a memory-centric agentic system for the full scientific research lifecycle. AutoSci is organized around four modules. SciMem provides schema-governed research memory, separating Long-Term Knowledge Memory for reusable scientific knowledge from Active Research Memory for project-level artifacts such as ideas, experiments, manuscripts, and reviews. SciFlow executes a five-stage lifecycle from literature understanding to rebuttal through a harness that controls state, context, verification, feedback, and orchestration. SciDAG augments difficult skills with DAG-shaped multi-agent operators and reusable stage-specific templates. SciEvolve converts feedback signals from users, experiments, reviews, and external environments into versioned updates to SciMem organization, SciFlow skills, and SciDAG templates. Together, these modules make AutoSci a persistent research environment that can execute, remember, and evolve across research projects. The code repository is available at https://github.com/skyllwt/AutoSci.

cs.AI↗

arXiv2026

Data-aware candidate selection in NL2SQL translation via small separating instances

We propose a data-aware candidate selection method for NL2SQL translation based on separating instances and provenance. We implement this approach and evaluate it against three natural baselines on a subset of BIRD-DEV. Experiments show that our method significantly outperforms baselines when only two or three candidates are given and no consistency score is available. The code of our prototype can be found at https://github.com/staskikotx/SISelection

cs.DB↗

arXiv2026

Memo-SQL: Structured Decomposition and Experience-Driven Self-Correction for Training-Free NL2SQL

Existing NL2SQL systems face two critical limitations: (1) they rely on in-context learning with only correct examples, overlooking the rich signal in historical error-fix pairs that could guide more robust self-correction; and (2) test-time scaling approaches often decompose questions arbitrarily, producing near-identical SQL candidates across runs and diminishing ensemble gains. Moreover, these methods suffer from a stark accuracy-efficiency trade-off: high performance demands excessive computation, while fast variants compromise quality. We present Memo-SQL, a training-free framework that addresses these issues through two simple ideas: structured decomposition and experience-aware self-correction. Instead of leaving decomposition to chance, we apply three clear strategies, entity-wise, hierarchical, and atomic sequential, to encourage diverse reasoning. For correction, we build a dynamic memory of both successful queries and historical error-fix pairs, and use retrieval-augmented prompting to bring relevant examples into context at inference time, no fine-tuning or external APIs required. On BIRD, Memo-SQL achieves 68.5% execution accuracy, setting a new state of the art among open, zero-fine-tuning methods, while using over 10 times fewer resources than prior TTS approaches.

cs.AI↗

arXiv2025

Deep learning models are vulnerable, but adversarial examples are even more vulnerable

Understanding intrinsic differences between adversarial examples and clean samples is key to enhancing DNN robustness and detection against adversarial attacks. This study first empirically finds that image-based adversarial examples are notably sensitive to occlusion. Controlled experiments on CIFAR-10 used nine canonical attacks (e.g., FGSM, PGD) to generate adversarial examples, paired with original samples for evaluation. We introduce Sliding Mask Confidence Entropy (SMCE) to quantify model confidence fluctuation under occlusion. Using 1800+ test images, SMCE calculations supported by Mask Entropy Field Maps and statistical distributions show adversarial examples have significantly higher confidence volatility under occlusion than originals. Based on this, we propose Sliding Window Mask-based Adversarial Example Detection (SWM-AED), which avoids catastrophic overfitting of conventional adversarial training. Evaluations across classifiers and attacks on CIFAR-10 demonstrate robust performance, with accuracy over 62% in most cases and up to 96.5%.

cs.CV↗

arXiv2025

A Survey of Data Agents: Emerging Paradigm or Overstated Hype?

The rapid advancement of large language models (LLMs) has spurred the emergence of data agents, autonomous systems designed to orchestrate Data + AI ecosystems for tackling complex data-related tasks. However, the term "data agent" currently suffers from terminological ambiguity and inconsistent adoption, conflating simple query responders with sophisticated autonomous architectures. This terminological ambiguity fosters mismatched user expectations, accountability challenges, and barriers to industry growth. Inspired by the SAE J3016 standard for driving automation, this survey introduces the first systematic hierarchical taxonomy for data agents, comprising six levels that delineate and trace progressive shifts in autonomy, from manual operations (L0) to a vision of generative, fully autonomous data agents (L5), thereby clarifying capability boundaries and responsibility allocation. Through this lens, we offer a structured review of existing research arranged by increasing autonomy, encompassing specialized data agents for data management, preparation, and analysis, alongside emerging efforts toward versatile, comprehensive systems with enhanced autonomy. We further analyze critical evolutionary leaps and technical gaps for advancing data agents, especially the ongoing L2-to-L3 transition, where data agents evolve from procedural execution to autonomous orchestration. Finally, we conclude with a forward-looking roadmap, envisioning proactive, generative data agents.

cs.DB↗

arXiv2025

WebRenderBench: Enhancing Web Interface Generation through Layout-Style Consistency and Reinforcement Learning

Automating the conversion of UI images into web code is a critical task for front-end development and rapid prototyping. Advances in multimodal large language models (MLLMs) have made WebUI-to-Code increasingly feasible, yet existing benchmarks remain limited in data diversity and evaluation reliability. To address these issues, we present WebRenderBench, a large-scale benchmark of 45.1k webpages collected from real-world portal sites, offering greater diversity, complexity, and realism than prior benchmarks. We further propose a novel evaluation metric that measures layout and style consistency from the final rendered pages. Unlike vision-based methods that rely on costly LLM reasoning or structure-based comparisons vulnerable to noise and asymmetry, our approach enables more efficient, objective, and reliable UI quality assessment. Finally, we introduce the Automated Layout and Style Inspection Agent (ALISA), which integrates this metric into reinforcement learning as a reward signal to enhance training on crawled asymmetric webpages. Experiments show that ALISA significantly boosts generation performance, achieving state-of-the-art results across multiple metrics.

cs.AI↗

arXiv2025

SAS-Bench: A Fine-Grained Benchmark for Evaluating Short Answer Scoring with Large Language Models

Subjective Answer Grading (SAG) plays a crucial role in education, standardized testing, and automated assessment systems, particularly for evaluating short-form responses in Short Answer Scoring (SAS). However, existing approaches often produce coarse-grained scores and lack detailed reasoning. Although large language models (LLMs) have demonstrated potential as zero-shot evaluators, they remain susceptible to bias, inconsistencies with human judgment, and limited transparency in scoring decisions. To overcome these limitations, we introduce SAS-Bench, a benchmark specifically designed for LLM-based SAS tasks. SAS-Bench provides fine-grained, step-wise scoring, expert-annotated error categories, and a diverse range of question types derived from real-world subject-specific exams. This benchmark facilitates detailed evaluation of model reasoning processes and explainability. We also release an open-source dataset containing 1,030 questions and 4,109 student responses, each annotated by domain experts. Furthermore, we conduct comprehensive experiments with various LLMs, identifying major challenges in scoring science-related questions and highlighting the effectiveness of few-shot prompting in improving scoring accuracy. Our work offers valuable insights into the development of more robust, fair, and educationally meaningful LLM-based evaluation systems.

cs.CL↗

arXiv2025

Alpha-SQL: Zero-Shot Text-to-SQL using Monte Carlo Tree Search

Text-to-SQL, which enables natural language interaction with databases, serves as a pivotal method across diverse industries. With new, more powerful large language models (LLMs) emerging every few months, fine-tuning has become incredibly costly, labor-intensive, and error-prone. As an alternative, zero-shot Text-to-SQL, which leverages the growing knowledge and reasoning capabilities encoded in LLMs without task-specific fine-tuning, presents a promising and more challenging direction. To address this challenge, we propose Alpha-SQL, a novel approach that leverages a Monte Carlo Tree Search (MCTS) framework to iteratively infer SQL construction actions based on partial reasoning states. To enhance the framework's reasoning capabilities, we introduce LLM-as-Action-Model to dynamically generate SQL construction actions during the MCTS process, steering the search toward more promising SQL queries. Moreover, Alpha-SQL employs a self-supervised reward function to evaluate the quality of candidate SQL queries, ensuring more accurate and efficient query generation. Experimental results show that Alpha-SQL achieves 69.7% execution accuracy on the BIRD development set, using a 32B open-source LLM without fine-tuning. Alpha-SQL outperforms the best previous zero-shot approach based on GPT-4o by 2.5% on the BIRD development set.

cs.DB↗

arXiv2022

Flow-Level Packet Loss Detection via Sketch Decomposition and Matrix Optimization

For cloud service providers, fine-grained packet loss detection across data centers is crucial in improving their service level and increasing business income. However, the inability to obtain sufficient measurements makes it difficult owing to the fundamental limit that the wide-area network links responsible for communication are not under their management. Moreover, millisecond-level delay jitter and clock synchronization errors in the WAN disable many tools that perform well in data center networks on this issue. Therefore, there is an urgent need to develop a new tool or method. In this work, we propose SketchDecomp, a novel loss detection method, from a mathematical perspective that has never been considered before. Its key is to decompose sketches upstream and downstream into several sub-sketches and builds a low-rank matrix optimization model to solve them. Extensive experiments on the test bed demonstrate its superiority.

cs.NI↗

arXiv2022

Multi-mode Tensor Train Factorization with Spatial-spectral Regularization for Remote Sensing Images Recovery

Tensor train (TT) factorization and corresponding TT rank, which can well express the low-rankness and mode correlations of higher-order tensors, have attracted much attention in recent years. However, TT factorization based methods are generally not sufficient to characterize low-rankness along each mode of third-order tensor. Inspired by this, we generalize the tensor train factorization to the mode-k tensor train factorization and introduce a corresponding multi-mode tensor train (MTT) rank. Then, we proposed a novel low-MTT-rank tensor completion model via multi-mode TT factorization and spatial-spectral smoothness regularization. To tackle the proposed model, we develop an efficient proximal alternating minimization (PAM) algorithm. Extensive numerical experiment results on visual data demonstrate that the proposed MTTD3R method outperforms compared methods in terms of visual and quantitative measures.

eess.IV↗

arXiv2021

"Sparse + Low-Rank'' Tensor Completion Approach for Recovering Images and Videos

Recovering color images and videos from highly undersampled data is a fundamental and challenging task in face recognition and computer vision. By the multi-dimensional nature of color images and videos, in this paper, we propose a novel tensor completion approach, which is able to efficiently explore the sparsity of tensor data under the discrete cosine transform (DCT). Specifically, we introduce two ``sparse + low-rank'' tensor completion models as well as two implementable algorithms for finding their solutions. The first one is a DCT-based sparse plus weighted nuclear norm induced low-rank minimization model. The second one is a DCT-based sparse plus $p$-shrinking mapping induced low-rank optimization model. Moreover, we accordingly propose two implementable augmented Lagrangian-based algorithms for solving the underlying optimization models. A series of numerical experiments including color image inpainting and video data recovery demonstrate that our proposed approach performs better than many existing state-of-the-art tensor completion methods, especially for the case when the ratio of missing data is high.

cs.CV↗

arXiv2021

Low-Rank and Sparse Enhanced Tucker Decomposition for Tensor Completion

Tensor completion refers to the task of estimating the missing data from an incomplete measurement or observation, which is a core problem frequently arising from the areas of big data analysis, computer vision, and network engineering. Due to the multidimensional nature of high-order tensors, the matrix approaches, e.g., matrix factorization and direct matricization of tensors, are often not ideal for tensor completion and recovery. In this paper, we introduce a unified low-rank and sparse enhanced Tucker decomposition model for tensor completion. Our model possesses a sparse regularization term to promote a sparse core tensor of the Tucker decomposition, which is beneficial for tensor data compression. Moreover, we enforce low-rank regularization terms on factor matrices of the Tucker decomposition for inducing the low-rankness of the tensor with a cheap computational cost. Numerically, we propose a customized ADMM with enough easy subproblems to solve the underlying model. It is remarkable that our model is able to deal with different types of real-world data sets, since it exploits the potential periodicity and inherent correlation properties appeared in tensors. A series of computational experiments on real-world data sets, including internet traffic data sets, color images, and face recognition, demonstrate that our model performs better than many existing state-of-the-art matricization and tensorization approaches in terms of achieving higher recovery accuracy.

cs.LG↗

arXiv2021

A Convergent Semi-Proximal Alternating Direction Method of Multipliers for Recovering Internet Traffics from Link Measurements

It is challenging to recover the large-scale internet traffic data purely from the link measurements. With the rapid growth of the problem scale, it will be extremely difficult to sustain the recovery accuracy and the computational cost. In this work, we propose a new Sparsity Low-Rank Recovery (SLRR) model and its Schur Complement Based semi-proximal Alternating Direction Method of Multipliers (SCB-spADMM) solver. Our approach distinguishes itself mainly for the following two aspects. First, we fully exploit the spatial low-rank property and the sparsity of traffic data, which are barely considered in the literature. Our model can be divided into a series of subproblems, which only relate to the traffics in a certain individual time interval. Thus, the model scale is significantly reduced. Second, we establish a globally convergent ADMM-type algorithm inspired by [Li et al., Math. Program., 155(2016)] to solve the SLRR model. In each iteration, all the intermediate variables' optimums can be calculated analytically, which makes the algorithm fast and accurate. Besides, due to the separability of the SLRR model, it is possible to design a parallel algorithm to further reduce computational time. According to the numerical results on the classic datasets Abilene and GEANT, our method achieves the best accuracy with a low computational cost. Moreover, in our newly released large-scale Huawei Origin-Destination (HOD) network traffics, our method perfectly reaches the seconds-level feedback, which meets the essential requirement for practical scenarios.

math.OC↗

arXiv2020

A Parallelizable Method for Missing Internet Traffic Tensor Data

Recovery of internet network traffic data from incomplete observed data is an important issue in internet network engineering and management. In this paper, by fully combining the temporal stability and periodicity features in internet traffic data, a new separable optimization model for internet data recovery is proposed, which is based upon the t-product and the rapid discrete Fourier transform of tensors. Moreover, by using generalized inverse matrices, an easy-to-operate and effective algorithm is proposed. In theory, we prove that under suitable conditions, every accumulation point of the sequence generated by the proposed algorithm is a stationary point of the established model. Numerical simulation results carried on the widely used real-world internet network datasets, show good performance of the proposed method. In the case of moderate sampling rates, the proposed method works very well, its effect is better than that of some existing internet traffic data recovery methods in the literature. The separable structural features presented in the optimization model provide the possibility to design more efficient parallel algorithms.

math.NA↗

arXiv2011

Efficient Continual Top-$k$ Keyword Search in Relational Databases

Keyword search in relational databases has been widely studied in recent years because it does not require users neither to master a certain structured query language nor to know the complex underlying data schemas. Most of existing methods focus on answering snapshot keyword queries in static databases. In practice, however, databases are updated frequently, and users may have long-term interests on specific topics. To deal with such a situation, it is necessary to build effective and efficient facility in database systems to support continual keyword queries evaluation. In this paper, we propose an efficient method for continual keyword queries answering over relational databases. The proposed method consists of two core algorithms. The first one computes a set of potential top-$k$ results by evaluating the ranges of the future relevance score for every query result and create a light-weight state for each keyword query. The second one uses these states to maintain the top-$k$ results of keyword queries when the database is continually growing. Experimental results validate the effectiveness and efficiency of the proposed method.

cs.DB↗