arXiv subjects

Meihao Fan

Publications and source records attributed to Meihao Fan.

9 recordsLinked to original sources

arXiv2026

SkillAdam: Stable and Efficient Skill Evolution for Agents

Agent skills provide a lightweight way to equip frozen language-model agents with domain knowledge and procedural guidance, yet obtaining high-quality skills remains costly and difficult to scale. Expert-written skills require substantial human effort. Recent skill self-evolution methods automate an iterative loop that uses execution feedback to revise skills, but their heuristic update strategies often yield unstable optimization and low iteration efficiency. We identify two challenges in realizing stable and efficient skill self-evolution. Direction Stability requires effective corrections to accumulate rather than be overwritten by iteration-local feedback. Update Adaptivity requires the scope of each revision to reflect the consistency of recent case-level improvements. We introduce SkillAdam, an Adam-inspired framework for optimizing discrete and non-differentiable skill documents. As a functional analogue of Adam's first moment, an optimization memory records identified problems and the outcomes of prior solution attempts to stabilize the update direction. As a functional analogue of Adam's second moment, a volatility-driven edit budget tracks the history-weighted variation of recent case-level improvements and adaptively controls the update magnitude. Across seven benchmarks that span short- and long-horizon tasks, SkillAdam achieves state-of-the-art performance with more stable optimization dynamics. It also obtains stronger skills with substantially fewer optimization iterations and lower cost than prior methods. Code repository: https://github.com/ruc-datalab/SkillAdam

cs.AI↗

arXiv2026

CODA-BENCH: Can Code Agents Handle Data-Intensive Tasks?

Advanced agents are increasingly demonstrating the potential to operate as autonomous engineers, creating a growing demand for evaluation benchmarks that capture the complexity of real-world development. Such environments typically involve both complex code and large-scale data (i.e., file system). However, existing benchmarks usually evaluate code-centric or data-centric capabilities in isolation, leaving a clear gap with real development scenarios. In this paper, we bridge this gap by introducing CODA-BENCH, the first benchmark to jointly evaluate code and data intelligence in a data-intensive environment. We construct a data-intensive Linux sandbox based on the Kaggle ecosystem (containing hundreds of datasets), where agents must actively explore complex file hierarchies to identify relevant resources and generate code for data-driven analytical tasks. CODA-BENCH comprises 1,009 tasks spanning 31 communities, with each task environment containing an average of 980 files, simulating realistic data scale and noise. Evaluations of advanced agents reveal that even top-performing systems struggle to effectively integrate data discovery with code execution, achieving a success rate of only 61.1%. These results highlight a substantial gap in current agentic capabilities for data-intensive tasks and point to promising directions for future research.

cs.AI↗

arXiv2026

TACO: A Benchmark for Open-Domain Text-to-SQL with Ambiguous and Cross-Database Queries

Text-to-SQL aims to translate natural language questions into executable SQL queries over structured databases. Existing benchmarks mainly focus on closed-domain settings with predefined database schemas and well-specified questions, but they fall short in addressing the challenges of open-domain scenarios, such as ambiguous questions, unspecified databases, and cross-database querying. To bridge this gap, we introduce TACO, a benchmark for open-domain Text-to-SQL with Ambiguous and Cross-database queries. TACO consists of 1,500 real-world Text-to-SQL examples based on a smart city data service and 13,000 high-quality synthetic examples generated based on large-scale open data portals, covering diverse domains such as transportation, healthcare, and finance. To construct the synthetic examples, we develop an effective data synthesis pipeline that preserves the complexity of real-world queries. To demonstrate the utility of TACO, we introduce a baseline TACO-SQL composed of question rewriting, table linking, and query planning, to illustrate the challenges posed by TACO and to better understand the limitations of existing Text-to-SQL approaches. Extensive experiments on TACO using a variety of recent Text-to-SQL approaches show that, while TACO-SQL achieves the best results, a significant gap still remains between the existing approaches and human-written SQL. These findings highlight the difficulty of open-domain Text-to-SQL and position TACO as a valuable benchmark to drive future research.

cs.DB↗

arXiv2026

DeepPrep: An LLM-Powered Agentic System for Autonomous Data Preparation

Data preparation, which aims to transform heterogeneous and noisy raw tables into analysis-ready data, remains a major bottleneck in data science. Recent approaches leverage large language models (LLMs) to automate data preparation from natural language specifications. However, existing LLM-powered methods either make decisions without grounding in intermediate execution results, or rely on linear interaction processes that offer limited support for revising earlier decisions. To address these limitations, we propose DeepPrep, an LLM-powered agentic system for autonomous data preparation. DeepPrep constructs data preparation pipelines through iterative, execution-grounded interaction with an environment that materializes intermediate table states and returns runtime feedback. To overcome the limitations of linear interaction, DeepPrep organizes pipeline construction with tree-based agentic reasoning, enabling structured exploration and non-local revision based on execution feedback. To enable effective learning of such behaviors, we propose a progressive agentic training framework, together with data synthesis that supplies diverse and complex ADP tasks. Extensive experiments show that DeepPrep achieves data preparation accuracy comparable to strong closed-source models (e.g., GPT-5) while incurring 15x lower inference cost, while establishing state-of-the-art performance among open-source baselines and generalizing effectively across diverse datasets.

cs.DB↗

arXiv2025

DeepAnalyze: Agentic Large Language Models for Autonomous Data Science

Autonomous data science, from raw data sources to analyst-grade deep research reports, has been a long-standing challenge, and is now becoming feasible with the emergence of powerful large language models (LLMs). Recent workflow-based data agents have shown promising results on specific data tasks but remain fundamentally limited in achieving fully autonomous data science due to their reliance on predefined workflows. In this paper, we introduce DeepAnalyze-8B, the first agentic LLM designed for autonomous data science, capable of automatically completing the end-toend pipeline from data sources to analyst-grade deep research reports. To tackle high-complexity data science tasks, we propose a curriculum-based agentic training paradigm that emulates the learning trajectory of human data scientists, enabling LLMs to progressively acquire and integrate multiple capabilities in real-world environments. We also introduce a data-grounded trajectory synthesis framework that constructs high-quality training data. Through agentic training, DeepAnalyze learns to perform a broad spectrum of data tasks, ranging from data question answering and specialized analytical tasks to open-ended data research. Experiments demonstrate that, with only 8B parameters, DeepAnalyze outperforms previous workflow-based agents built on most advanced proprietary LLMs. The model, code, and training data of DeepAnalyze are open-sourced, paving the way toward autonomous data science.

cs.AI↗

arXiv2025

Reward-SQL: Boosting Text-to-SQL via Stepwise Execution-Aware Reasoning and Process-Supervised Rewards

Recent advances in large language models (LLMs) trained with reinforcement learning (RL) have improved Text-to-SQL performance. However, RL-based approaches still struggle with complex queries due to two key limitations: insufficient stepwise execution-aware reasoning grounded in database feedback, and the lack of process-level rewards for guiding reasoning optimization. To address these issues, we propose CoCTE, a divide-and-conquer and execution-aware reasoning framework that progressively composes SQL queries through intermediate view validation and structured Common Table Expressions (CTEs), improving both accuracy and interpretability. To realize a CoCTE reasoning process, we develop Reward-SQL, a unified approach with three stages: (1) model initialization, which equips LLMs with structured CoCTE reasoning capabilities; (2) process reward design, which delivers fine-grained, execution-aware supervision; and (3) process-supervised RL and inference, which integrates process rewards into training and guides the inference stage by process rewards. This paper addresses the core challenges in Reward-SQL and makes the following contributions. We introduce a process reward model (PRM) that combines execution-aware trajectory scoring with entropy-based step weighting, providing dense and interpretable supervision across reasoning steps. We integrate PRM into both RL training and inference stages, stabilizing optimization and improving trajectory exploration with process-level signals. Experiments show that Reward-SQL significantly outperforms baselines with comparable model sizes, and exhibits strong cross-domain generalization.

cs.CL↗

arXiv2024

AutoPrep: Natural Language Question-Aware Data Preparation with a Multi-Agent Framework

Answering natural language (NL) questions about tables, known as Tabular Question Answering (TQA), is crucial because it allows users to quickly and efficiently extract meaningful insights from structured data, effectively bridging the gap between human language and machine-readable formats. Many of these tables are derived from web sources or real-world scenarios, which require meticulous data preparation (or data prep) to ensure accurate responses. However, preparing such tables for NL questions introduces new requirements that extend beyond traditional data preparation. This question-ware data preparation involves specific tasks such as column derivation and filtering tailored to particular questions, as well as question-aware value normalization or conversion, highlighting the need for a more nuanced approach in this context. Because each of the above tasks is unique, a single model (or agent) may not perform effectively across all scenarios. In this paper, we propose AutoPrep, a large language model (LLM)-based multiagent framework that leverages the strengths of multiple agents, each specialized in a certain type of data prep, ensuring more accurate and contextually relevant responses. Given an NL question over a table, AutoPrep performs data prep through three key components. Planner: Determines a logical plan, outlining a sequence of high-level operations. Programmer: Translates this logical plan into a physical plan by generating the corresponding low-level code. Executor: Executes the generated code to process the table. To support this multi-agent framework, we design a novel Chain-ofClauses reasoning mechanism for high-level operation suggestion, and a tool-augmented method for low-level code generation.

cs.CL↗

arXiv2023

Cost-Effective In-Context Learning for Entity Resolution: A Design Space Exploration

Entity resolution (ER) is an important data integration task with a wide spectrum of applications. The state-of-the-art solutions on ER rely on pre-trained language models (PLMs), which require fine-tuning on a lot of labeled matching/non-matching entity pairs. Recently, large languages models (LLMs), such as GPT-4, have shown the ability to perform many tasks without tuning model parameters, which is known as in-context learning (ICL) that facilitates effective learning from a few labeled input context demonstrations. However, existing ICL approaches to ER typically necessitate providing a task description and a set of demonstrations for each entity pair and thus have limitations on the monetary cost of interfacing LLMs. To address the problem, in this paper, we provide a comprehensive study to investigate how to develop a cost-effective batch prompting approach to ER. We introduce a framework BATCHER consisting of demonstration selection and question batching and explore different design choices that support batch prompting for ER. We also devise a covering-based demonstration selection strategy that achieves an effective balance between matching accuracy and monetary cost. We conduct a thorough evaluation to explore the design space and evaluate our proposed strategies. Through extensive experiments, we find that batch prompting is very cost-effective for ER, compared with not only PLM-based methods fine-tuned with extensive labeled data but also LLM-based methods with manually designed prompting. We also provide guidance for selecting appropriate design choices for batch prompting.

cs.CL↗

arXiv2021

Few-shot Multi-hop Question Answering over Knowledge Base

KBQA is a task that requires to answer questions by using semantic structured information in knowledge base. Previous work in this area has been restricted due to the lack of large semantic parsing dataset and the exponential growth of searching space with the increasing hops of relation paths. In this paper, we propose an efficient pipeline method equipped with a pre-trained language model. By adopting Beam Search algorithm, the searching space will not be restricted in subgraph of 3 hops. Besides, we propose a data generation strategy, which enables our model to generalize well from few training samples. We evaluate our model on an open-domain complex Chinese Question Answering task CCKS2019 and achieve F1-score of 62.55% on the test dataset. In addition, in order to test the few-shot learning capability of our model, we ramdomly select 10% of the primary data to train our model, the result shows that our model can still achieves F1-score of 58.54%, which verifies the capability of our model to process KBQA task and the advantage in few-shot Learning.

cs.CL↗