About Me
I’m a third-year Computer Science student at the College of Computer Science and Technology, Xi’an Jiaotong University, expected to earn my B.S. in Engineering in fall 2027. My research interests primarily focus on Large Language Model Agents in Domain-specific Scenarios, RL, and Computer Use (CLI, GUI). Contact me at jiayuw794@gmail.com.
Roadmap
|
Efficient Agent Scaffolding & Harness Paper: Designing Domain-Specific Agents via Hierarchical Task Abstraction Mechanism(2025.11 arxiv) Project: Antlet-CLI Blog: Why Does Claude Code Work So Well on Non‑Coding Tasks? |
Agent-based Process Automation & Dev Tools Project: Contribot Project: ShipNuts(ongoing) Project: research-navigator(ongoing) Project: Earth-Insights WeChat Official Account Agent (地球洞察微信公众号) Project: dl-reproduce Blog: After a Vibe Coding Interview |
|
| Efficient Autonomous Large Language Model Agent and AI System | ||
|
LLM agent in domain-specific application Paper: Designing Domain-Specific Agents via Hierarchical Task Abstraction Mechanism(2025.11 arxiv) Project: EarthAgent |
Computer Use Agent Project: AARR Series Paper: Act As a Real Researcher: A Suite of Benchmarks Evaluating Frontier LLMs and Agentic Harnesses in Research Lifecycle(2026.05 arxiv) Project: Auto-Cursor Blog: Why I'm Still Bullish on GUI Agents in 2026 |
|
News
- 2026.05: 🎉🎉 We launch AARR Series(Link) and publish AARRI-bench(Link)!
- 2026.04: 🎉🎉 “SPD-Faith Bench” have been accepted in ACL 2026(Link)!
- 2026.03: 🎉🎉 Contribot(Link) releases!
- 2026.03: 🎉🎉 I launch an open source rust project: Antlet(Link)!
- 2026.03: 🎉🎉 dl-reproduce skill has released!
- 2026.01: 🎉🎉 I attend the semi-final competition of AI Agent 2025(Link) held by (Chinese Association for Artificial Intelligence)(CAAI) in Shenzhen, China!
- 2025.12: 🎉🎉 I get the merit award(￥1500) working with my teammates zepeng and weijiang in Harmony System Control Agent Competition(Link) held by Nanjing University and Huawei!
- 2025.11: 🎉🎉 The number of subscribers of the Account has reached 500!
- 2025.11: 🎉🎉 GeoPlan-bench Releases!
- 2025.11: 🎉🎉 Earth-Insights WeChat Official Account’s Paper-Deep-Dive Feature Releases!
- 2025.10: 🎉🎉 I attend “Vibe Coding hackthon” held by WaytoAGI(Link) and give a talk on GUI-Agents and their potential in the future!
- 2025.10: 🎉🎉 Auto-Cursor Releases!
- 2025.10: 🎉🎉 Earth-Insights Account has started cross-posting on RedNote!
- 2025.10: 🎉🎉 Earth-Insights WeChat Official Account’s first Semi-Weekly-Report Release!
- 2025.09: 🎉🎉 I attend Cursor Meetup Xi’an!
- 2025.09: 🎉🎉 EarthAgent succeeds in MVP test by AI Agent 2025 Committee!
- 2025.08: 🎉🎉 I participate in the “First National College Student Artificial Intelligence Security Competition” (Link) held at Beijing University of Posts and Telecommunications and win the first prize!
- 2025.01: 🎉🎉 I attend Khalifa University Winter Youth Camp (Link) in Abu Dhabi, UAE!
Publications
Act As a Real Researcher: A Suite of Benchmarks Evaluating Frontier LLMs and Agentic Harnesses in Research Lifecycle
| Paper | Code | Project |
Jiayu Wang, Weijiang Lv, Bowen Fu, Jing Fu, Jiayi Song, Lingyu Zhang, Lanxuan Xue, Luodi Chen, Zepeng Xin, Kaiyu Li, Xiangyong Cao
- Launch the AARR benchmark series, a novel framework for evaluating the capabilities of LLM agents in authentic research scenarios.
- Propose AARRI-Bench, the inaugural benchmark in this series, which comprises tasks designed to simulate real research intern activities.
- Conduct extensive experiments across frontier models and agentic systems, providing a comprehensive analysis of their current capabilities and limitations.
GeoFaith: A Spatio-Temporal Dual View of Faithful Chain-of-Thought
Weijiang Lv, Wentong Zhao, Jiayu Wang, Yuhao Wu, Jiaheng Wei, Xiaobo Xia
- Introduce a spatio-temporal perspective on Chain-of-Thought faithfulness, characterizing reasoning via latent geometric structure and entropy dynamics.
- Develop a scalable bootstrapping pipeline that expands step-level faithfulness annotations from 1k to 20k examples across multiple domains, enabling large-scale process supervision.
- Design a faithfulness-aware reinforcement learning framework that jointly optimizes outcome correctness, process faithfulness, and trajectory consistency, achieving superior detection and generation performance while producing shorter, more interpretable reasoning chains.
Designing Domain-Specific Agents via Hierarchical Task Abstraction Mechanism
| Paper | Code | Project |
Kaiyu Li, Jiayu Wang, Zhi Wang, Hui Qiao, Weizhan Zhang, Deyu Meng, Xiangyong Cao
- We introduce a novel agent design framework centered on a Hierarchical Task Abstraction Mechanism (HTAM).
- We instantiate this framework as EarthAgent, a multi-agent system tailored for complex geospatial analysis. To evaluate such complex planning capabilities, we build GeoPlan-bench, a comprehensive benchmark of realistic, multi-step geospatial planning tasks. It is accompanied by a suite of carefully designed metrics to evaluate tool selection, path similarity, and logical completeness.
- Experiments show that EarthAgent substantially outperforms a range of established single- and multi-agent systems.
SPD-Faith Bench: Diagnosing and Improving Faithfulness in Chain-of-Thought for Multimodal Large Language Models
| Paper | Code | Data |
Weijiang Lv，Yaoxuan Feng，Xiaobo Xia，Jiayu Wang，Yan Jing，Wenchao Chen，Bo Chen
- We introduce SPD-Faith Bench, a diagnostic benchmark based on fine-grained image difference reasoning that enforces explicit visual comparison.
- Evaluations on state-of-the-art MLLMs reveal two systematic failure modes, perceptual blindness and perception-reasoning dissociation.
- We propose SAGE, a train-free visual evidence-calibrated framework that improves visual routing and aligns reasoning with perception. Results highlight the importance of explicitly evaluating faithfulness beyond response correctness
DescribeEarth: Describe Anything for Remote Sensing Images
| Paper | Code | Dataset | Benchmark |
Kaiyu Li, Zixuan Jiang, Xiangyong Cao, Jiayu Wang, Yuchen Xiao, Deyu Meng, Zhi Wang
- We propose Geo-DLC, a novel task of object-level fine-grained image captioning for remote sensing.
- We construct DE-Dataset, a large-scale dataset containing 25 categories and 261,806 annotated instances with detailed descriptions of object attributes, relationships, and contexts.
- Furthermore, we introduce DE-Benchmark, an LLM-assisted question-answering based evaluation suite designed to systematically measure model capabilities on the Geo-DLC task.
- We also present DescribeEarth, a Multi-modal Large Language Model (MLLM) architecture explicitly designed for Geo-DLC.
- Our DescribeEarth model consistently outperforms state-of-the-art general MLLMs on DE-Benchmark, demonstrating superior factual accuracy, descriptive richness, and grammatical soundness, particularly in capturing intrinsic object features and surrounding environmental attributes across simple, complex, and even out-of-distribution remote sensing scenarios.
Projects (selected)
AARR Series (link) (May 2026)
Evaluating how well LLM agents close the gap with human researchers across the full research lifecycle. Not just executing code — testing the cognitive gaps that still separate frontier agents from human researchers.
Antlet-CLI (on going) (link) (May 2026)
A nano coding agent built with Rust, supporting memory management, and cron-triggered tasks. It has extremely fast startup speed and minimal memory footprint, and natively supports running a massive number of agent instances simultaneously.
Contribot (link) (March 2026)
A system that automatically contributes to GitHub open-source repositories using tons of Claude Code as the reasoning engine. It continuously monitors your target repos, analyzes issues and codebase, then submits PRs under your own GitHub account.
Each Claude Code instance will be assigned a target repo. It will analyze this repo systematically, scan the current issues and PRs and find a gap to work on. The instance itself will test and review in local working directory, making sure the PRs are high quality and valuable.
It runs days and nights without taking a nap. The results are not AI slops, but actually beyond my expectation. Many PRs made by Contribot are highly valuable and have been merged into the official repo.
Auto-Cursor (link) (Oct 2025)
Auto-Cursor is THE FIRST (to the best of my knowledge) GUI-native orchestration layer that pilots the Cursor IDE like a human operator. By combining large language models, visual grounding, and deterministic automation, the project explores how agents can build software without being confined to command-line tooling.
Why Through GUI?
Human-parity reach: Command-line automation is capped by the APIs that tools expose. A GUI agent, however, can click, type, drag, and navigate any surface that a human can. This dramatically widens the solution space—if a person can operate it, an agent can learn to operate it too, opening the door to automating entire product lifecycles.
Grounded perception is ready: Domain-specific MLLMs now recognize icons, layouts, and context with far higher reliability. The bottleneck has shifted from perception to orchestration. Auto-Cursor focuses on that orchestration layer—sequencing vision, language, and action—to unlock richer, end-to-end workflows.
Standing on the shoulders of the ecosystem: GUI-first control leverages advances in agents, LLMs, GPU-accelerated rendering, and even display hardware. We treat the modern desktop as a programmable environment, turning existing tools into improvable building blocks instead of rewriting them.
Vision
Build a resilient, self-improving system that can iterate on its own behaviors, learn from failures, and adapt to different project constraints.
Provide tangible GUI agent scenarios that inspire new ideas for downstream industries—design, ops, education, assistive tech, and beyond.
Stimulate thinking on AI safety and software design, showing how oversight, logging, and guardrails can coexist with highly capable automation.
EarthAgent (link) (July 2025)
EarthAgent is a groundbreaking general AI agent for the remote sensing field, dedicated to making complex and high-threshold geospatial analysis more accessible and automated. It allows users to drive a fully automated workflow that integrates multimodal data acquisition, intelligent interpretation, and deep reasoning through simple natural language conversations. Whether with text or image inputs, EarthAgent can autonomously plan and execute tasks, reducing traditional manual analysis processes that used to take days to just minutes. It has attracted 300+ likes on RedNote. link
Earth-Insights WeChat Official Account Agent (地球洞察微信公众号) (Oct 2025)
This is an automated system that fetches the latest papers in the fields of remote sensing and deep learning. Through sophisticated design and arrangement, it utilizes document analysis agents and various document analysis tools to achieve fully automated analysis and summarization. Currently, it consists of two modules: the Semi-Weekly Report and the Paper Deep Dive. As of December 16, the official account has published a total of 100+ blogs, covering 400+ papers, accumulating over 6,000 reads, and attracting more than 700 followers, providing the community with convenient access to the latest information. Semi-Weekly Report Demo: link Paper Deep Dive Demo: link
Blogs
Is It Possible That Future Software Is Just Pure Text?
Software with Only Natural Language
Read more →
Why Does Claude Code Work So Well on Non‑Coding Tasks?
Generalization of Agent Scaffolding
Read more →
Why I'm Still Bullish on GUI Agents in 2026
The Rise, Current Situation, and Future of GUI-Agent
Read more →
Contribution
Open Source Contributions
Published MCP Servers
A set of public MCP (Model Context Protocol) servers I built and published, available on Dedalus Labs Marketplace and my GitHub. Each connects an AI agent to a popular SaaS platform — detailed information can be found in the related repositories on my GitHub.
