Reinforcement Learning for Vision-Language Models

Microsoft AI • Redmond, WA

2d • $119,800 - $304,200

About The Position

Microsoft AI is seeking a Member of Technical Staff to advance frontier vision-language models through reinforcement learning. As part of the Large Scale Reinforcement Learning team, this role focuses on helping models understand visual information, reason from visual evidence, and respond accurately and usefully to complex multimodal questions. Areas of work include visual perception and grounding, document and chart understanding, image-to-code generation, and spatial and physical reasoning from visual inputs. The goal is to build capabilities that generalize to unfamiliar problems and real-world use, rather than merely improve benchmark scores.

This hands-on research and engineering role involves owning projects across the research loop: identifying capability gaps, developing training tasks and data, designing reliable reward signals, running large-scale experiments, and improving models based on results. The successful candidate will collaborate with researchers and engineers across reinforcement learning, multimodal modeling, data, infrastructure, evaluation, and safety.

Requirements

• Bachelor's degree in Computer Science, Machine Learning, Mathematics, or a related technical field and 4+ years of relevant research or engineering experience; or a master's degree in a related field and 2+ years of relevant experience; or a PhD in a related field.
• Experience developing, training, or evaluating deep learning models, with demonstrated depth in at least one of model post-training, reinforcement learning, or vision-language modeling.
• Ability to formulate hypotheses, design informative experiments, analyze results, and debug model or training behavior.
• Ability to take end-to-end ownership of technical projects and collaborate effectively across research and engineering disciplines.
• Proficient programming skills in Python and PyTorch or a similar framework.

Nice To Haves

• Bachelor's degree in Computer Science, Machine Learning, Mathematics, or a related technical field and 8+ years of relevant experience; or a master's degree and 6+ years; or a PhD and 4+ years of relevant research or engineering experience.
• Experience training models via reinforcement learning across distributed environments.
• Demonstrated ability to build models for complex visual reasoning tasks such as visual grounding, document and chart understanding, image-to-code generation, and spatial or physical reasoning.
• Experience developing synthetic data, reward models, automated verifiers, or model-based evaluation systems, including assessing reliability and failure modes.
• Meaningful research or engineering contributions demonstrated through publications, open-source work, deployed systems, or other substantial technical projects.

Responsibilities

• Advance visual understanding and multimodal reasoning through reinforcement learning by translating capability gaps into research hypotheses and measurable improvements.
• Develop targeted training tasks and data, including synthetic data, to address capability gaps.
• Design and validate rewards, verifiers, rubrics, and model-based judges for verifiable and open-ended visual tasks, and detect and mitigate reward hacking.
• Build evaluations and conduct controlled experiments, ablations, and detailed failure analysis.
• Collaborate with scaling, multimodal, data, and safety teams to integrate improvements while preserving broader model capabilities.

What This Job Offers

• Job Type: Full-time
• Career Level: Mid Level