Research Engineer, Frontier Capabilities

Lila Sciences Lila Sciences · AI Frontier · One Charles Park, Cambridge, MA · AI

Research Engineer focused on training LLMs for long-horizon scientific discovery tasks, spanning the post-training stack from SFT to asynchronous RL on agentic harnesses. The role involves designing, building, and optimizing systems for scaling post-training, sharpening reasoning, and enabling compute-intensive agentic-harness training. Specific work streams include GPU optimization, stack and infrastructure development, model experimentation, evaluations and benchmarks, and agentic capabilities research.

What you'd actually do

  1. Design, build, and optimize systems to push this frontier: scaling post-training, sharpening reasoning, and unlocking compute-intensive agentic-harness training.
  2. Maximize hardware utilization across 100B+ parameter asynchronous RL training runs.
  3. Own the post-training infrastructure end-to-end — supervised fine-tuning, asynchronous RL with tool integration, and data pipelines.
  4. Lead experimentation on reasoning model development, including mixture-of-experts stabilization, curriculum design, and synthetic reasoning trace generation.
  5. Design and build best-in-class scientific agentic benchmarks and harnesses, along with the dashboards and leaderboards that inform every training decision.
  6. Train models capable of planning, exploration, and tool use over extended horizons.

Skills

Required

  • Strong software engineering skills in Python
  • Experience with distributed ML training frameworks (Megatron-LM, TorchTitan, DeepSpeed, Ray)
  • Understanding of large-scale model training techniques for 100B+ models
  • Experience with cloud or HPC environment
  • Ability to communicate technical results to internal and external stakeholders

Nice to have

  • C++/CUDA
  • Prior work with large scale scientific datasets or domain-specific modeling
  • Contributions to open-source ML frameworks
  • Experience with RL post-training (RLHF, GRPO, tool-augmented RL)
  • Experience training MoE architectures

What the JD emphasized

  • training LLMs to run long-horizon scientific discovery tasks
  • full post-training stack
  • asynchronous RL on agentic harnesses
  • scaling post-training
  • sharpening reasoning
  • compute-intensive agentic-harness training
  • Maximize hardware utilization across 100B+ parameter asynchronous RL training runs
  • Own the post-training infrastructure end-to-end
  • supervised fine-tuning
  • asynchronous RL with tool integration
  • data pipelines
  • modular, reproducible workflows with single-command execution
  • composable pipelines spanning Data, SFT, and RL stages
  • productionize novel algorithms to run at scale
  • reasoning model development
  • mixture-of-experts stabilization
  • curriculum design
  • synthetic reasoning trace generation
  • experimental design and tracking
  • scientific agentic benchmarks and harnesses
  • agentic benchmarks and harnesses
  • planning, exploration, and tool use over extended horizons
  • RL at scale with tool-calling
  • subgoal decomposition
  • shared memory/skills across trials
  • scientific agent capabilities
  • Experience with distributed ML training frameworks
  • Understanding of large-scale model training techniques for 100B+ models
  • Experience with RL post-training (RLHF, GRPO, tool-augmented RL)

Other signals

  • training LLMs
  • long-horizon scientific discovery tasks
  • full post-training stack
  • SFT to asynchronous RL
  • agentic harnesses
  • plan, use tools, learn from experience
  • scaling post-training
  • sharpening reasoning
  • compute-intensive agentic-harness training
  • design, build, and optimize systems
  • GPU Optimization & Training Performance
  • Maximize hardware utilization
  • custom kernel development
  • Stack & Infrastructure
  • Own the post-training infrastructure end-to-end
  • supervised fine-tuning
  • asynchronous RL with tool integration
  • data pipelines
  • modular, reproducible workflows
  • composable pipelines spanning Data, SFT, and RL stages
  • productionize novel algorithms
  • Model Experimentation
  • reasoning model development
  • mixture-of-experts stabilization
  • curriculum design
  • synthetic reasoning trace generation
  • experimental design and tracking
  • Evaluations & Benchmarks
  • scientific agentic benchmarks and harnesses
  • dashboards and leaderboards
  • agentic benchmarks and harnesses
  • Agentic Capabilities & Frontier Research
  • planning, exploration, and tool use over extended horizons
  • RL at scale with tool-calling
  • subgoal decomposition
  • shared memory/skills across trials
  • scientific agent capabilities
  • Strong software engineering skills in Python
  • C++/CUDA a plus
  • distributed ML training frameworks
  • Megatron-LM
  • TorchTitan
  • DeepSpeed
  • Ray
  • large-scale model training techniques for 100B+ models
  • cloud or HPC environment
  • communicate technical results
  • Prior work with large scale scientific datasets
  • domain-specific modeling
  • Contributions to open-source ML frameworks
  • RL post-training (RLHF, GRPO, tool-augmented RL)
  • training MoE architectures