Algorithm Researcher — Agentic Vision

About the role

Tencent's Technology Engineering Group (TEG) is hiring an Algorithm Researcher focused on agentic vision. The role takes ownership of full-stack algorithm research for post-training of Vision-Language Models (VLMs), covering fine-tuning, visual perception, reliability and user interaction scenarios.

Key responsibilities

• Lead post-training research for VLMs, exploring supervised fine-tuning (SFT) and reinforcement learning alignment methods such as RLHF, PPO and GRPO.
• Optimize low-level visual perception, fine-grained understanding and spatial awareness, and explore visual agent solutions such as thinking with images.
• Reduce hallucinations in multimodal large models by investigating their causes, building detection pipelines, refining data cleaning and developing reinforcement learning-based removal solutions.
• Align and optimize models for real-world user scenarios, improving reasoning, instruction following and detail perception in multi-turn image-text interactions, and set up an online badcase monitoring and data feedback loop.

Required profile

• Master's degree or above in Computer Science, Software Engineering, Artificial Intelligence or a related field.
• Extensive research and hands-on experience in at least one post-training domain such as LLM alignment, RLHF, reward modeling, personalized LLMs or memory/RAG mechanisms.
• Experience in user profiling, recommendation systems integrated with large models, or long-context/long-term memory model optimization.
• Experience training or fine-tuning models at tens-of-billions to hundreds-of-billions parameter scale.
• Publications at top-tier conferences (NeurIPS, ICLR, ICML, ACL, EMNLP, etc.) or significant contributions to high-impact open-source projects.

Required skills

• Deep understanding of the Transformer architecture and large language model training principles.
• Proficiency in Python and familiarity with PyTorch and TensorFlow.
• Hands-on experience with distributed training and large-scale training/inference frameworks such as Megatron-LM, DeepSpeed and vLLM.