Jinxiang Meng孟金翔
I am currently interning at ByteDance Seed.
I am a Computer Science master’s student at the University of Chinese Academy of Sciences (UCASNJ), trained at the Institute of Automation, Chinese Academy of Sciences, and advised by Prof. Kang Liu.
My research primarily focuses on general agents. Recently, I have been exploring personal agents and proactive agents.
Seeking full-time opportunities in general agents. Let’s connect.
News
- Seed2.1 was released. I contributed to improving its general agent capabilities.
- Our paper "DV-World: Benchmarking Data Visualization Agents in Real-World Scenarios" was accepted to ICML 2026.
- Our paper "GATE: Graph-based Adaptive Tool Evolution Across Diverse Tasks" was accepted to ACL 2026.
Earlier updates
- Seed2.0 was released. My main contribution focused on improving the Seed model's subjective data analysis capability through data construction and post-training data optimization for open-ended data analysis tasks.
- Our paper "DAComp: Benchmarking Data Agents across the Full Data Intelligence Lifecycle" was accepted to ICLR 2026.
- Our paper "TaREx: Reinforcement Learning for Code-Driven Table Reasoning" was accepted as an Oral at AAAI 2026.
- We released the paper, code, and datasets for "Reasoning-Table: Exploring Reinforcement Learning for Table Reasoning".
- We released the paper and code for "GATE: Graph-based Adaptive Tool Evolution Across Diverse Tasks".
Publications
* Equal contributionPapers & preprints
DAComp: Benchmarking Data Agents across the Full Data Intelligence Lifecycle
TaREx: Reinforcement Learning for Code-Driven Table Reasoning
GATE: Graph-based Adaptive Tool Evolution Across Diverse Tasks
SWE-Touch: Benchmarking Coding Agents When Users Touch the Code
StartupBench: Benchmarking General-Purpose Agents on Market-Validated End-to-End Workflows
Reasoning-Table: Exploring Reinforcement Learning for Table Reasoning
Technical reports
Seed2.1 Model Card: Agentic Intelligence for Productivity
I contributed to improving Seed2.1’s general agent capabilities.
Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity
Under review
Not All Failures Are Equal: Pattern-Guided Fine-Tuning for Improving LLM Agents
SheetForge: Scaling Spreadsheet Agents in Executable Environments
DA-Omni: Scaling Training for Full-Stack Data Agents
VerifierArena: Benchmarking and Analyzing Universal Judges for Unverifiable Agentic Tasks
Experience
ByteDance Seed
Research Intern
Education
University of Chinese Academy of Sciences
Master’s student in Computer Science · UCASNJ / CASIA
Advisor: Prof. Kang Liu · GPA: 3.82/4.00 · Rank: 1/35
Tianjin University
Undergraduate visiting student · Information Management and Information Systems
Academic Service
Reviewer: AAAI 2027 · NeurIPS 2026 · ICLR 2027