Job description

The Meta Reality Labs Research Team brings together researchers, developers, and engineers to create the future of AR, VR, and contextual AI. The Surreal Vision group at Reality Labs Research is seeking outstanding Research Scientist Interns to advance posttraining of vision-language and omni models. As an intern, you will develop reinforcement learning (RL) methods that improve multimodal reasoning, temporal understanding, and grounded decision-making. Applications include procedural and real-time task assistants for future contextual AI devices. Your research project may cover or relate to the following topics: - Developing RL posttraining algorithms for VLMs and omni models, including on-policy optimization and learning from expert feedback. - Designing reward models and verifiable rewards for multimodal reasoning, temporal grounding, and calibrated decisions. - Combining supervised fine-tuning, preference learning, and RL to improve data efficiency, training stability, and generalization. - Posttraining with streaming video, speech, and environmental audio to improve reasoning. - Building evaluations and ablations that measure grounded model improvements and identify reward hacking and failures under ambiguous evidence. Our internships are twelve (12) to twenty-four (24) weeks long and we have flexible start dates. Plan and execute innovative research to advance the state of the art in multimodal posttraining and reinforcement learning. Collaborate with researchers to develop training pipelines, reward functions, experiments, and prototype systems. Benchmark RL methods against supervised fine-tuning and preference-optimization baselines, and analyze model behavior and generalization. Publish research findings at leading conferences and present research outcomes to internal and external audiences. Currently has, or is in the process of obtaining, a PhD in machine learning, computer vision, natural language processing, speech and audio processing, reinforcement learning, or a related field Proven research contributions demonstrated by grants, fellowships, patents, or publications at leading workshops or conferences such as NeurIPS, ICML, CVPR, ICLR, ICCV, ECCV, ACL, EMNLP, NAACL, ICASSP, and Interspeech Knowledge of deep learning and multimodal foundation models, with hands-on experience in VLM or omni model posttraining, particularly reinforcement learning Experience implementing deep learning algorithms, training models, and designing benchmarks and evaluations Programming experience in Python and deep learning frameworks such as PyTorch Experience working and communicating across functions in a team environment Must obtain work authorization in the country of employment at the time of hire, and maintain ongoing work authorization during employment Experience with RL posttraining methods, reward modeling, and/or preference optimization Experience with distributed training, rollout generation, and scalable evaluation of multimodal foundation models Research experience in video understanding, temporal reasoning, and/or audio-visual learning Demonstrated research experience through internships, work experience, open source contributions, and/or publications Intent to return to a degree program after completion of the internship

Requirements

Must-haves

Your research project may cover or relate to the following topics:
Developing RL posttraining algorithms for VLMs and omni models, including on-policy optimization and learning from expert feedback.
Designing reward models and verifiable rewards for multimodal reasoning, temporal grounding, and calibrated decisions.
Combining supervised fine-tuning, preference learning, and RL to improve data efficiency, training stability, and generalization.
Posttraining with streaming video, speech, and environmental audio to improve reasoning.
Building evaluations and ablations that measure grounded model improvements and identify reward hacking and failures under ambiguous evidence.
Our internships are twelve (12) to twenty-four (24) weeks long and we have flexible start dates.

Nice-to-haves

None flagged.

Application prep

Use the role’s language and application details while tailoring your materials.

Keywords to mirror

aimachine learningpython

Platform: meta
Effort: medium

Benefits

Benefits detected in the listing.

Vision