Mengsha Liu

I am an LLM algorithm engineer at ByteDance in Singapore, working on agentic search: models that decompose and rewrite their own queries and orchestrate search tools over long, multi-turn trajectories.

I work on both halves of that loop — agentic RL with verifiable and rubric-based rewards for long-horizon tool use, and the harness around it: action-space design, context management, and the data engine that synthesises trajectories and mines rollouts. Evaluation should score the process, not only the answer.

Previously: core search at Huawei, open-world game agents at Tencent AI Lab, multimodal document and chart understanding at Alibaba DAMO Academy. M.Sc. from NTU, B.Eng. from Sun Yat-sen University with Prof. Ying Shen. Open to collaboration.

Mengsha Liu

News

  • 2026.08 Declare, Compile, Look accepted to the ECCV 2026 Multimodal Digital Agents workshop, selected for a talk.
  • 2026.05 Released a technical report on automating the e-commerce search relevance loop with agents.
  • 2025.07 Joined ByteDance in Singapore as an LLM algorithm engineer, working on agentic search.
  • 2025.06 LTCR published in Neural Computing and Applications.
  • 2025.06 Graduated with an M.Sc. in Signal Processing and Machine Learning from NTU Singapore.

Publications

Declare, Compile, Look pipeline — plan a typed FigureSpec, compile the geometry deterministically, emit figures, and repair by patching the spec

Declare, Compile, Look:
Coordinate-Free Layout Generation with Vision-in-the-Loop Repair

Guian Fang, Mengsha Liu, Mike Zheng Shou

ECCV Workshop (MDA), 2026Talk

Paper
Case-driven multi-agent framework: a User Agent mines bad cases with an Annotator, an Optimizer pipeline turns them into data and model fixes, and a shared harness supplies tools, skills and human interaction

A Case-Driven Multi-Agent Framework
for E-Commerce Search Relevance

Global E-Commerce Search Relevance Team, ByteDance

Technical Report, 2026

Replaces the human roles in the relevance loop with agents: a User Agent surfaces bad cases through conversation, an Annotator Agent labels them over multiple turns, and an Optimizer Agent diagnoses and resolves them — all on a shared retrieval-augmented harness.

Report
Salience-aware fake news detection model — two attention-based LSTM branches, the first re-weighting its input through a Gaussian-filtered attention layer before the features are aggregated

LTCR:
Long-Text Chinese Rumor Detection Dataset

Mengsha Liu, Ziyang Ma, Guian Fang, Ying Shen

Neural Computing and Applications, 2025

A long-form Chinese misinformation benchmark built around COVID-19 rumours, released with a salience-aware detection model.

Paper arXiv
ChartThinker overview — context retrieval feeds a chain-of-thought pipeline that turns a chart into a grounded summary

ChartThinker:
A Contextual Chain-of-Thought Approach to Optimized Chart Summarization

Mengsha Liu, Daoyuan Chen, Yaliang Li, Guian Fang, Ying Shen

LREC-COLING, 2024

Released Chart-Sum-QA — a large-scale chart–caption and instruction dataset covering a wide range of topics and visual styles.

Paper arXiv Code Dataset

Experience

Large-model work in industry, across search, e-commerce and game environments.

  1. 2025

    ByteDance

    LLM Algorithm Engineer

    2025.07 — Present · Singapore

    • Search agents that plan retrieval, call search tools and rewrite their own queries over several turns, for long-tail queries a single pass cannot answer.
    • A daily pipeline that mines bad cases from the training set, synthesises harder ones, and corrects labels on similar queries already in the corpus.
  2. 2025

    Huawei

    Search Algorithm Intern

    2025.01 — 2025.05 · Singapore

    • Core search algorithms, including how chain-of-thought reasoning behaves inside a production search stack.
  3. 2024

    Tencent AI Lab

    Large-Model Algorithm Intern

    2024.03 — 2024.08 · Shenzhen

    • Generative agents for open worlds: planning and executing long, multi-step skills in a changing environment.
    • NPC dialogue and schedule generation for the Honor of Kings open-world project; the schedule generator was filed as a patent.
  4. 2023

    Alibaba DAMO Academy

    Large-Model Algorithm Intern

    2023.06 — 2023.12 · Shenzhen

    • Chain-of-thought information extraction from long documents and charts.
    • Multimodal instruction tuning on 600K image–text pairs and 8M instruction QA samples.

Education

  1. 2024

    Nanyang Technological University

    M.Sc.

    2024.08 — 2025.06 · Singapore

    Signal Processing and Machine Learning

  2. 2020

    Sun Yat-sen University

    B.Eng.

    2020.09 — 2024.06 · Guangzhou

    Intelligent Science and Technology · GPA top 1% of cohort

Patents & Awards

Patents & Software Copyright

Scholarships