Algorithm Engineer, Multimodal Agent Post-Training

What you’ll do

1. 01Build post-training data systems for multimodal agents to improve VLM capabilities. Develop and iterate on high-quality training data covering visual understanding, visual reasoning, GUI / Computer Use, and agent interaction trajectories.
2. 02Use model capability analysis and failure cases to design new training tasks, data distributions, and evaluation plans that improve model performance on complex multimodal tasks.
3. 03Build data and environments for Computer Use and Multimodal agent tasks, including interactive task design, environment development, trajectory data production, and data quality assessment.
4. 04Research and apply data strategies for post-training methods such as SFT, Preference Learning, and RL, exploring more effective ways to improve multimodal agent capabilities.
5. 05Track advances in VLMs, Multimodal agent systems, and Computer Use, and explore new data paradigms and evaluation methods based on business needs.

What we look for

1. 01Experience with VLMs, large multimodal models, or agent post-training, with an understanding of post-training workflows and how data improves model capabilities.
2. 02Strong data design and problem analysis skills, with the ability to design effective data tasks based on model performance and failure cases.
3. 03Familiarity with multimodal model techniques and practical experience in areas such as visual understanding, Grounding, multimodal reasoning, or GUI Agent systems.
4. 04Experience with Computer Use, GUI Agent systems, Web agent systems, agent evaluation, or interactive environments is preferred.
5. 05Proficiency in Python, with skills in model experimentation, data processing, and software development.
6. 06Experience with large-scale data construction, synthetic data, model evaluation, or benchmark design is preferred.

Nice to have

• Experience in post-training, data construction, or building evaluation systems for mainstream VLMs or Multimodal agent models.
• Project experience with Computer Use, GUI Agent systems, or Web agent systems.
• Experience with agent trajectories, environment development, reinforcement learning data, or reward design.
• Experience building benchmarks, evaluation datasets, or data quality assessment systems.
• Papers at leading conferences, open-source projects, or research experience with large models.