Your impact at LILA

We’re building a talent-dense, high-agency research team to develop the next generation of learning systems and reasoning algorithms for agentic LLMs. Our work sits at the intersection of large language models, post-training, and scientific reasoning, with the goal of enabling systems that learn from experience, reason effectively, and improve through interaction.

Scientific domains present a distinct set of challenges that make this problem uniquely hard. Feedback is sparse and delayed — experiments take days or weeks, not milliseconds. Ground truth is expensive or contested. Distribution shift is structural, as instruments, techniques, and knowledge bases evolve continuously. The hypothesis space is vast and reward signal is thin. Existing benchmark do not capture these nuances. The goal is to build systems that can operate effectively in this scientific regime.

This role spans a few complementary directions. Candidates are expected to bring deep expertise in one (ore more) of the following areas. In the event of cross-track expertise, please select the one you align to the most. Our interview process will be catered to verifying the chosen expertise area.

Expertise Area 1 — Agentic system building

Focus: Build systems that autonomously propose, execute, and verify scientific hypotheses over long time horizons.

• Create and analyze long-running auto-research systems that propose and verify hypotheses
• Design planning frameworks for agentic systems operating over long, sparse feedback loops
• Design memory architectures that allow agents to build and retrieve structured knowledge over time
• Explore algorithms in recursive self-improvement, multi-agent coordination, and continual learning

Expertise Area 2: Distillation

Focus: Translate strong inference-time behaviors and reasoning traces into efficient, trainable models.

• Develop distillation strategies from large or ensemble models into deployable systems
• Research methods for self-improvement, including iterative self-distillation and critique loops
• Investigate how to preserve generalization and reduce catastrophic forgetting through the distillation process

Expertise Area 3 — Scalable experience generation

Focus: Develop inference-time algorithms and synthetic data pipelines that generate high-quality training signal for scientific reasoning.

• Design and benchmark inference-time search, sampling, and verification strategies
• Propose new techniques in synthetic environment creation and curriculum learning
• Develop synthetic data generation strategies that capture high-quality scientific reasoning for agentic model training
• Measure the end-to-end impact of inference-time improvements on real scientific tasks

What you’ll need to succeed:

• An advanced degree in computer science, machine learning, or a related field, or or comparable experience
• Strong foundation in LLMs and empirical research
• Experience designing and executing rigorous ML experiments, including benchmarking and ablations
• Experience working with large-scale training or evaluation pipelines
• Ability to define and pursue research directions in open-ended, rapidly evolving spaces
• Strong collaboration and communication skills across research and engineering teams

Bonus points for:

• Experience with synthetic data generation, distillation, or self-improvement loops
• Familiarity with reinforcement learning (e.g., RLHF, on-policy methods)
• Experience with planning, search, or decision-making systems at scale
• Experience in building agentic systems with tool use, or multi-agent workflows
• Background in program synthesis, coding benchmarks, or long-horizon tasks
• Experience building evaluation frameworks or large-scale benchmarks

Scientific rigor & persistence:

• You take a principled approach to experimentation, with careful baselines, ablations, and evaluation design
• You are motivated by understanding why systems work, not just improving metrics
• You prioritize clarity, reproducibility, and intellectual honesty in research
• You are comfortable working through long, nonlinear iteration cycles
• You operate effectively in ambiguous, fast-evolving research environments

Your Impact at LILA

The AI Research team is tackling one of the most exciting, open problems in AI: training LLMs to run long-horizon scientific discovery tasks. Our approach spans the full post-training stack - from SFT to asynchronous RL on agentic harnesses - teaching models to plan, use tools, and learn from experience in domains where the ground truth isn't a preference label, but a scientific result.

We're rapidly growing our Research Engineering org and seeking talented engineers and ML practitioners across levels to design, build, and optimize systems to push this frontier: scaling post-training, sharpening reasoning, and unlocking compute-intensive agentic-harness training. This is a rare chance to join an early team with the autonomy, flexibility, and compute to tackle frontier science problems.

We operate with high agency, and a bias toward execution. Below are several focus areas within the team. We ask that candidates select the stream that best matches their experience and excitement.

Work Streams

Stream A: GPU Optimization & Training Performance

Maximize hardware utilization across 100B+ parameter asynchronous RL training runs. Responsibilities include profiling, performance optimization, custom kernel development, communication-computation overlap, and long-context throughput improvements. You set and maintain the performance baseline.

Stream B: Stack & Infrastructure

Own the post-training infrastructure end-to-end — supervised fine-tuning, asynchronous RL with tool integration, and data pipelines. Build modular, reproducible workflows with single-command execution. Manage upstream framework upgrades and deliver composable pipelines spanning Data, SFT, and RL stages. You work tightly with Research Scientists to develop and productionize novel algorithms to run at scale.

Stream C: Model Experimentation

Bring deep, hands-on experience training large language models. Lead experimentation on reasoning model development, including mixture-of-experts stabilization, curriculum design, and synthetic reasoning trace generation. You have a bias toward experimental design and tracking, and know how to prioritize runs that yield promising outcomes.

Stream D: Evaluations & Benchmarks

Design and build best-in-class scientific agentic benchmarks and harnesses, along with the dashboards and leaderboards that inform every training decision. You have experience working with well known public benchmarks and have spent time building bespoke agentic benchmarks and harnesses.

Stream E: Agentic Capabilities & Frontier Research

Train models capable of planning, exploration, and tool use over extended horizons. Advance the state of the art in RL at scale with tool-calling, subgoal decomposition, and shared memory/skills across trials to expand the frontier of scientific agent capabilities.

What You'll Need to Succeed

• Strong software engineering skills in Python; C++/CUDA a plus
• 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

Bonus Points For

• 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

Location

San Francisco, CA or Cambridge, MA (Remote, Hybrid, and On-Site available depending on team needs).