Vision-Language AI Scientist: Post-Training & RLHF

Tencent; Singapore; On-site; SGD 180,000 - 280,000; Full time

Job summary

Tencent’s Technology Engineering Group is seeking a senior researcher to own full-stack R&D for post-training of vision-language models. You will explore supervised fine-tuning and reinforcement learning alignment methods, aiming to push the capability ceiling of vision models.

You will advance low-level visual perception, improve multimodal reliability, and design evaluation pipelines for real-world image-text interactions, including online feedback loops to refine outcomes.

Qualifications

• Master's degree or higher in Computer Science, Artificial Intelligence or related field.
• Deep understanding of Transformer architectures and LLM training principles.
• Hands-on experience in at least one post-training domain (LLM Alignment, RLHF, Reward Modeling, Personalization, Memory/RAG).
• Experience with user profiling and recommender systems integrated with large models or long-context memory optimization.
• Strong Python skills and familiarity with PyTorch and TensorFlow.

Responsibilities

• Lead end-to-end post-training R&D for vision-language models including SFT and RL-based algorithms.
• Advance visual foundation perception and spatial reasoning for multimodal models.
• Develop and refine hallucination-detection pipelines and reinforcement learning-based removal methods.
• Align models with real-world user scenarios and multi-turn image-text interactions.

Skills

Python

Transformer

LLM training

RLHF

PPO

DeepSpeed

Megatron-LM

vLLM

PyTorch

TensorFlow

Education

Master's degree or above in Computer Science, AI, or related field

Tools

Megatron-LM

DeepSpeed

vLLM

HuggingFace

PyTorch

TensorFlow

Job description

Tencent’s Technology Engineering Group is seeking a senior researcher to own full-stack R&D for post-training of vision-language models. You will explore supervised fine-tuning and reinforcement learning alignment methods, aiming to push the capability ceiling of vision models.

You will advance low-level visual perception, improve multimodal reliability, and design evaluation pipelines for real-world image-text interactions, including online feedback loops to refine outcomes.