Research Scientist - Large Language Model

About Luma AI

Luma’s mission is to build AGI. We believe that intelligence emerges from large-scale foundation models that can reason, plan, and communicate with depth and precision. Language models are central to this vision — serving as the backbone for reasoning, world modeling, and interaction.

To advance this mission, we build and operate the full stack end-to-end, spanning foundation models, large-scale training infrastructure, inference systems, and real-world products. This tight integration allows us to push research forward while shipping impactful systems at scale. Backed by a recent $900M Series C and our partnership with Humain to build a 2 GW compute supercluster (Project Halo), we are scaling the next generation of frontier language models.

Where You Come In

This is a rare opportunity to help define the future of large-scale language models. You will work across the entire lifecycle of model development — from large-scale pre-training, to targeted mid-training, to post-training alignment and capability refinement.

You will operate at the frontier of scaling laws, reasoning, and alignment, directly shaping how foundation models learn, generalize, and behave in real-world deployments.

What You’ll Do

This role spans both the “science” and “engineering” dimensions of research — two aspects that are equally important.

You will work across modeling, data, systems, and evaluation.

Modeling

• Architect and scale large autoregressive language models.
• Design improved pre-training objectives to enhance reasoning, knowledge retention, and compositional generalization.
• Develop mid-training strategies such as continued pre-training, domain adaptation, curriculum learning, and synthetic data integration.
• Advance post-training techniques, including instruction tuning, preference optimization, reinforcement learning, distillation, and inference-time compute scaling.
• Study and improve long-context modeling, planning depth, and multi-step reasoning behavior.

Data

• Curate and construct massive, high-quality text corpora for pre-training.
• Design synthetic data pipelines for reasoning, tool use, mathematics, coding, and structured problem solving.
• Develop filtering, mixture weighting, and curriculum strategies that shape emergent capabilities.
• Formulate new tasks that improve coherence, logical consistency, factual grounding, and robustness.

Systems

• Train frontier-scale language models across large GPU clusters.
• Optimize distributed training (data, tensor, pipeline parallelism), mixed precision, and memory efficiency.
• Build infrastructure for large-scale experimentation, ablations, and reproducibility.
• Improve inference efficiency and support scalable deployment.

Evaluation

• Define and build evaluation frameworks for language intelligence, including:

Multi-step reasoning and mathematical problem solving

Coding and structured generation

Knowledge grounding and factuality

Planning and agentic behavior

Instruction following and alignment

• Track capability development across pre-training, mid-training, and post-training.
• Close the loop between evaluation signals and data/model improvements.

Who You Are

• Strong foundation in machine learning and large language models.
• Deep understanding of autoregressive transformers and large-scale training dynamics.
• Experience with pre-training large models and/or post-training techniques such as instruction tuning, RLHF, preference optimization, or distillation.
• Hands-on experience with PyTorch and distributed training at scale.
• Comfortable operating across research and production environments.

What Sets You Apart (Bonus Points)

• Experience training frontier-scale language models from scratch.
• Research contributions in scaling laws, reasoning, alignment, or inference-time compute.
• Experience designing large-scale synthetic reasoning data.
• Expertise in long-context modeling or structured reasoning systems.
• Experience optimizing models for real-world deployment constraints.

Your application are reviewed by real people.