Senior Research Scientist: Post-Training RL for Multimodal Medical AI

Epsilon Health; San Francisco (CA); On-site; USD 180,000 - 240,000; Full time

Job summary

Epsilon Health is transforming medical imaging with frontier AI. We seek a Research Scientist to own post-training, reward modeling, and reinforcement learning for radiology reports. You will work on multimodal models with clinical rigor, leveraging large-scale medical imaging data and cutting-edge RL techniques.

Join a team advancing state-of-the-art grounded report generation and inference-time strategies, while publishing and shaping deployment practices in healthcare AI.

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 GRPO/DAPO/GSPO/CISPO with multi-reward objectives
• Reinforcement learning with verifiable rewards, including noisy, sparse, or learned reward signals
• Reward model training: outcome and process supervision
• Preference optimization methods (DPO, IPO, ORPO, KTO) and RLHF
• Inference-time compute scaling: best-of-N sampling and verifier-guided decoding
• Practical experience diagnosing reward hacking and over-optimization
• Track record of implementing complex models from research to production
• Proficiency in PyTorch or JAX, with experience training large models on multi-GPU
• Experience with RL infrastructure (rollout generation) and frameworks such as verl, TRL, or OpenRLHF
• Autoregressive language modeling and instruction tuning
• Strong software engineering and production-quality coding

Responsibilities

• Design reinforcement learning with verifiable rewards for report generation in radiology workflows
• Extend RL to unverifiable/noisy objectives using learned reward models and radiologist feedback (RLHF)
• Run GRPO-family algorithms with multi-reward objectives and diagnose reward hacking
• Train explicit reward models with multimodal conditioning on images
• Train chain-of-thought reasoning over image regions with grounding
• Develop multimodal tool use with credit assignment across trajectories
• Create inference-time methods like best-of-N and grounding-aware decoding
• Tune output to institutional reporting conventions
• Stay current with reinforcement learning, reward modeling, and multimodal post-training
• Publish findings and establish best practices for post-training medical VLMs

Skills

reinforcement learning

post-training

multimodal ML

vision-language models

PyTorch

JAX

distributed systems

reinforcement learning infrastructure

Tools

vLLM

SGLang

TRL

OpenRLHF

GPUs

Job description

Epsilon Health is transforming medical imaging with frontier AI. We seek a Research Scientist to own post-training, reward modeling, and reinforcement learning for radiology reports. You will work on multimodal models with clinical rigor, leveraging large-scale medical imaging data and cutting-edge RL techniques.

Join a team advancing state-of-the-art grounded report generation and inference-time strategies, while publishing and shaping deployment practices in healthcare AI.