Agentic Vision & Multimodal AI Research Intern

Tencent; Singapore; On-site; SGD 140,000 - 210,000; Full time

Job summary

Tencent is seeking a researcher to advance Agentic Vision for multimodal foundation models in Singapore. You will focus on visual reasoning post-training, integrate perception, planning, tool execution, and environment feedback to enable dynamic reasoning.

Responsibilities include developing post-training methods, building data synthesis pipelines, and creating interactive sandboxes for tool use and evaluation. A strong background in deep learning and PyTorch is required.

Qualifications

• Master’s degree or higher in CS, AI, CV, ML, or a related field.
• Strong foundations in deep learning and algorithm implementation.
• Proficiency in PyTorch and distributed training.
• Experience with multimodal foundation models and visual reasoning.
• Familiarity with post-training methods (SFT, PPO/GRPO, OPD, RLVR).
• Experience with OpenRLHF, Megatron, or related frameworks.

Responsibilities

• Conduct R&D on Agentic Vision for Hunyuan multimodal foundation models, focusing on visual reasoning post-training.
• Advance visual reasoning and spatial intelligence, including 3D understanding and complex visual logic.
• Develop multimodal post-training methods (SFT, PPO/GRPO, OPD) and verifiers/reward models.
• Build verifiable data synthesis pipelines using rendering engines, symbolic systems, and code sandboxes.
• Create interactive sandboxes supporting code execution, visual tools, and evaluation protocols.

Skills

PyTorch proficiency

Distributed training experience

Multimodal research

Visual reasoning

Reinforcement learning

Education

Master’s degree or higher in CS/AI/CV/ML or related field

Tools

Megatron

OpenRLHF

Verl/OpenRLHF

Job description

Tencent is seeking a researcher to advance Agentic Vision for multimodal foundation models in Singapore. You will focus on visual reasoning post-training, integrate perception, planning, tool execution, and environment feedback to enable dynamic reasoning.

Responsibilities include developing post-training methods, building data synthesis pipelines, and creating interactive sandboxes for tool use and evaluation. A strong background in deep learning and PyTorch is required.