Responsibilities

We are looking for an exceptional post-training research scientist to develop methodologies across environment and state construction, data generation and mixing, training, and evaluation. Working closely with other researchers, you will help shape the research agenda and translate successful methods into production. Specifically, you will:
Develop post-training methods for forecasting and reasoning. Design and execute supervised fine-tuning, reinforcement learning, on-policy distillation, and related methods to improve accuracy and calibration.
Build environments and task distributions. Construct point-in-time environments combining historical observations, documents, prior predictions, and tools. Define tasks that require information gathering, belief updating, and forecasting across multiple horizons.
Develop reward modeling and credit assignment methods, combining verifiable intermediate outcomes, expert supervision, preference data, and delayed feedback.
Build efficient, reliable training systems. Collaborate with engineers to develop synchronous and asynchronous learning pipelines on multi-node GPU clusters.
Establish rigorous evaluations that measure forecasting accuracy, uncertainty calibration, reasoning quality, robustness, and downstream utility. Design temporal holdouts and controlled ablations that distinguish generalization from leakage, distribution shifts, and evaluator noise.
Productionize research. Collaborate with quantitative researchers to identify valuable model capabilities and deploy successful methods into reliable production systems.
Bachelor’s degree or higher in a quantitative or technical discipline, or equivalent practical experience developing advanced learning systems.
Demonstrated expertise in large language model (LLM) post-training, reinforcement learning, environment design, or preference learning. Experience working in non-fully verifiable domains is preferred.
Hands-on experience with distributed GPU training, including performance optimization, numerical stability, checkpointing, and reproducibility.
Strong implementation skills in Python and PyTorch or JAX for research and production model training.
Experience using distributed training and inference frameworks, including FSDP, DeepSpeed, SGLang, Megatron, verl, or Slime.

Post-Training Research Scientist - Forecasting and Reasoning at point72 · LumiJobs

Post-Training Research Scientist - Forecasting and Reasoning

Responsibilities

Qualification

Required

Benefits

Company

point72

More roles at point72