Role OverviewWe're seeking a
Research Scientist with deep expertise in
post-training and reinforcement learning to join our
ML Research team. You'll be at the forefront of developing and deploying state-of-the-art multimodal models for clinical use in radiology settings. This role owns every stage after pretraining: supervised fine-tuning, reward modeling, reinforcement learning against verifiable and learned reward signals, reasoning and tool-use training, and inference-time strategy. You'll work with one of the largest and most diverse medical imaging datasets in the industry, advancing the state-of-the-art in grounded report generation, reward design, and inference-time reasoning while maintaining the clinical rigor required for healthcare deployment.
Key Responsibilities- Design reinforcement learning with verifiable rewards for report generation, including clinical label and entity-relation matching, grounding IoU, measurement accuracy, and reporting schema compliance.
- Extend reinforcement learning to unverifiable and noisy objectives such as report quality and clinical usefulness, using learned reward models and radiologist feedback pipelines (RLHF) built on expert preferences and report edits.
- Run GRPO-family algorithms with complex multi-reward objectives, tuning reward composition and diagnosing reward hacking, entropy collapse, and diversity loss.
- Train explicit reward models, including multimodal reward models conditioned on the image, with both outcome and process supervision.
- Train chain-of-thought reasoning over image regions, including evidence localization and verification loops that keep reasoning grounded in the image rather than in language priors.
- Train multimodal tool use - windowing, zoom and crop, detector and segmentation calls, prior study retrieval - with credit assignment across multi-turn trajectories.
- Develop inference-time methods including best-of-N sampling against reward models and grounding-aware decoding, and distill the resulting gains back into the policy.
- Tune output stylization to institutional reporting conventions, keeping style rewards separated from clinical content rewards.
- Stay current with cutting-edge research in reinforcement learning, reward modeling, and multimodal post-training.
- Drive research and technical excellence through conference publications and technical blog posts, establishing best practices for post-training medical VLMs at scale.
Qualifications- 6+ years of academia/industry experience in reinforcement learning, post-training, or multimodal machine learning
- Deep expertise in post-training large language or vision-language models (e.g., Qwen-VL, InternVL, LLaVA, or similar architectures)
- Strong foundation in modern post-training and reinforcement learning techniques including:
- Group-relative policy optimization and its successors (GRPO, DAPO, GSPO, CISPO) with multi-reward objectives
- Reinforcement learning with verifiable rewards, and with noisy, sparse, or learned reward signals
- Reward model training: pairwise and generative reward models, outcome and process supervision
- Preference optimization methods (DPO, IPO, ORPO, KTO) and RLHF
- Inference-time compute scaling, including best-of-N sampling and verifier-guided decoding
- Practical experience diagnosing and mitigating reward hacking and reward over-optimization
- Track record of implementing complex models from research papers and adapting them to new domains
- Proficiency in PyTorch or JAX, with experience training large models on multi-GPU/distributed systems
- Experience with reinforcement learning infrastructure at scale, including rollout generation (vLLM, SGLang) and frameworks such as verl, TRL, or OpenRLHF
- Experience with autoregressive language modeling and instruction tuning
- Strong software engineering skills and ability to write production-quality code
Preferred Qualifications- Publications at top-tier conferences (NeurIPS, ICML, ICLR, CVPR, ACL, EMNLP, MICCAI)
- Hands-on experience with medical imaging applications, particularly radiology report generation
- Experience with agentic or multi-turn reinforcement learning, including credit assignment over tool-use trajectories
- Experience with grounded generation tasks (visual grounding, referring expression comprehension)
- Knowledge of evaluation methodologies for long-form generation, including factuality assessment and hallucination detection
- Experience mitigating catastrophic forgetting of supervised capabilities during reinforcement learning
- Familiarity with clinical NLP and medical knowledge representation
- Experience with model interpretability, explainability, and uncertainty quantification in safety-critical applications
