Shaojie Zhang

Research profile

Research summary

Shaojie Zhang is a researcher at Xiaomi (China) working on multimodal AI agents, with an Evolve catalog profile centered on two 2026 arXiv papers: MaP, which unifies omni-dimensional GUI agent navigation via masked trajectory prediction, and NTEP-R, which adds necessary tool-evidence path rewards to reinforcement learning for agentic vision-language models. The bibliometric record, resolved from OpenAlex, attributes 5 works in 2026 — the only year with recorded counts — and splits the field distribution evenly (0.5/0.5) between AI (GUI agents / vision-language models) and machine learning (multimodal AI agents). OpenAlex topic records reach further, led by Software Testing and Debugging Techniques (2 papers) alongside topic modeling, neural network applications, recommender systems, and reinforcement learning in robotics, suggesting a broader applied-ML topic footprint. Collaboration is the strongest scored dimension at 62, resting on 22 distinct coauthors and a stable frequent-coauthor core including Pei Fu and Yan Zhang; on both catalog papers Zhang holds a middle, non-corresponding position within large industry author teams. Activity scores 52 on the strength of the recent burst, while reproducibility scores 20 because both papers are rated low repro-readiness — MaP lacks significance tests and NTEP-R's edge over a reproduced baseline is a thin 2.03 points. Output and impact remain unscored given an h-index of 1, an i10-index of 0, and no multi-year publication counts.

Research keywords

Highlights

• MaP: unifying GUI agent navigation with masked trajectory prediction

  In this 2026 arXiv paper (middle author), MaP reformulates omni-dimensional GUI agent navigation as masked trajectory prediction, unifying tasks that mixture training handles separately. It reports consistent 2-3 point gains over mixture training across 5 benchmarks and uses LoRA adapters to ease gradient conflict in multi-task learning; Evolve notes the missing significance tests.

• NTEP-R: process rewards for necessary tool-evidence paths

  This 2026 arXiv paper adds necessary tool-evidence path rewards to reinforcement learning for agentic vision-language models, sharpening credit assignment for tool calls and delivering clear efficiency gains. Evolve cautions that the margin over one reproduced baseline is a thin 2.03 points and that rewards rely on a judge.

• A coherent thread: supervision at the trajectory and evidence level

  Across both 2026 papers the same idea recurs — supervise the agent's intermediate structure (trajectories for GUI navigation, evidence paths for tool use) rather than outcomes alone, implemented within LoRA-based and GRPO-based training pipelines for vision-language models.

Timeline

1. 2026Co-authored 'Towards Omni-dimensional GUI Agent Navigation with Masked Trajectory Prediction' (arXiv), contributing to MaP's masked trajectory prediction framework for GUI agents, in a middle author position.
2. 2026Co-authored 'Making Every Tool Call Count: Necessary Tool-Evidence Path Rewards for Agentic Vision-Language Models' (arXiv), the NTEP-R process-reward method for VLM agent reinforcement learning.
3. 2026OpenAlex counts-by-year attributes 5 works to this identity in 2026 — the only publication year recorded — matching recent3yWorks of 5.
4. 2026Topic profile spans five OpenAlex areas, led by Software Testing and Debugging Techniques (2), alongside topic modeling, neural network applications, recommender systems, and reinforcement learning in robotics.

Papers in the Evolve catalog

• 2026-09-22

  Towards Omni-dimensional GUI Agent Navigation with Masked Trajectory Prediction

• 2026-09-03

  Making Every Tool Call Count: Necessary Tool-Evidence Path Rewards for Agentic Vision-Language Models

Frequent co-authors