Research Scientist Graduate- (Foundation Model, Vision and Language) - 2025 Start (PhD)

Responsibilities

Welcome to the Doubao-Vision team, where we spearhead multi-modality foundation models on visual understanding and visual generation. Our mission is to solve the visual intelligence problem for AI. We conduct cutting-edge research on areas like vision and language, large vision models, and generative foundation models. The team is a mix of experienced research scientists and engineers, aiming to advance the research boundaries in foundation models and apply our technologies to our rich application scenarios, whereas a feedback loop is created to help further improve our foundation technologies. Join us in shaping the future of AI technologies and revolutionizing our product experience for global users. We are looking for talented individuals to join our team in 2025. As a graduate, you will get unparalleled opportunities for you to kickstart your career, pursue bold ideas and explore limitless growth opportunities. Co-create a future driven by your inspiration with Bytedance. Successful candidates must be able to commit to an onboarding date by end of year 2025. We will prioritize candidates who are able to commit to these start dates. Please state your availability and graduation date clearly in your resume. Applications will be reviewed on a rolling basis. We encourage you to apply early. Candidates can apply for a maximum of TWO positions and will be considered for jobs in the order you applied for. The application limit is applicable to Bytedance and its affiliates' jobs globally. Responsibilities - Conduct cutting-edge research and development in computer vision and natural language processing, especially in the areas of multi-modality, vision and language, etc. - Enhance multimodal understanding and reasoning (images and videos etc), throughout the entire development process, encompassing data acquisition, model evaluation, pre-training, SFT, reward modeling, and reinforcement learning, to bolster overall performance. - Synthesize large-scale, high-quality multi-modal data through methods such as rewriting, augmentation, and generation to improve the abilities of foundation models in various stages (pretraining, SFT, RLHF). - Investigate and implement robust evaluation methodologies to assess model performance at various stages (ranging from covering diverse multimodal skills to improving user preference alignment), unravel the underlying mechanisms and sources of their abilities, and utilize this understanding to drive model improvements.

Qualifications

Minimum Qualifications: - Research and engineering experience in one or more areas of computer vision and natural language processing, including but not limited to: Experience in multi-modal understanding, vision and language, such as multimodal pre-training, visual instruction tuning, alignment learning, and other related topics. - Work with very large-scale datasets, and build very large-scale datasets to scale up foundation models. - Experience with language models and apply them in various downstream tasks. - Highly competent in algorithms and programming; Strong coding skills in Python and popular deep learning frameworks. - Work and collaborate well with team members. - Ability to work independently; Strong communication skills. Preferred Qualifications: - Candidates with publications in top-tier venues such as CVPR, ECCV, ICCV, NeurIPS, ICLR, ICML, EMNLP, ACL, NAACL, etc - Candidates with impactful open-source projects on GitHub and a demonstrated engineering ability to quickly solve new challenges.