Vision-Language AI Researcher: LLM Alignment & Perception

Origen Tech; Singapore; On-site; SGD 120,000 - 190,000; Full time

Job summary

Origen Tech seeks a research scientist to lead post-training R&D for vision-language models. You will own end-to-end algorithm development for post-training, exploring SFT and RLHF, PPO, GRPO, and expanding the capabilities of vision foundation models.

You will advance visual perception and multimodal reasoning, reduce model hallucinations with detection pipelines and data cleaning, and improve real-world user interactions with robust evaluation in multi-turn image-text scenarios.

Qualifications

• Master's degree or higher in Computer Science, Software Engineering, AI, or related field.
• Strong understanding of Transformer architectures and LLM training principles.
• Experience in post-training domains including LLM Alignment, RLHF, and memory/RAG.
• Experience with large-scale model training and inference.

Responsibilities

• Own the full-stack algorithm R&D for post-training of vision-language models.
• Explore SFT and reinforcement learning alignment methods (RLHF, PPO, GRPO).
• Improve low-level visual perception and spatial understanding.
• Reduce hallucinations and build robust hallucination detection pipelines.
• Enhance multi-turn image-text interaction and real-world user scenarios.

Skills

Transformer

LLM Training

RLHF

Reward Modeling

Memory/RAG

Python

PyTorch

TensorFlow

Megatron-LM

DeepSpeed

vLLM

Publications

HuggingFace

Teamwork

Education

Master's degree or higher in CS/AI

Tools

Megatron-LM

DeepSpeed

vLLM

HuggingFace

Job description

Origen Tech seeks a research scientist to lead post-training R&D for vision-language models. You will own end-to-end algorithm development for post-training, exploring SFT and RLHF, PPO, GRPO, and expanding the capabilities of vision foundation models.

You will advance visual perception and multimodal reasoning, reduce model hallucinations with detection pipelines and data cleaning, and improve real-world user interactions with robust evaluation in multi-turn image-text scenarios.