We are seeking a Senior AI Researcher to lead post-training evaluation, red-teaming, and reinforcement learning (RL) gym audits on open-weight models. The ideal candidate will establish rigorous benchmarking methodologies, evaluate large language models (LLMs) against complex threats like Indirect Prompt Injections (IPI), and construct post-training evaluation pipelines that accurately measure realistic frontier-level security capabilities. Key Responsibilities RL Post-Training & Benchmarking
* Execute post-training runs (e.g. GRPO) using mainstream open-weight generalist models against security-focused RL environments, targeting threat vectors like Indirect Prompt Injection (IPI).
* Reward Diagnostics & Trace Analysis - Analyze live loss curves and rollout traces to identify reward hacking, lazy policy convergence, and flawed or over/under-specified verifiers. Task & Environment Auditing
* Review tasks and multi-turn environments (including tool use, web navigation, and computer use) for realism, threat model accuracy, data distribution, and dataset balance. Performance Reporting (Gym Cards)
* Generate comprehensive evaluation cards detailing hill-climbing performance uplift across checkpoints, failure modes, tokens/turns per rollout, and task-level success rates. Integration & Orchestration
* Integrate dockerized environments (e.g., Harbor format) into internal training frameworks, optimizing reset/statefulness semantics, concurrency, and throughput ceilings.
Hybrid:
Yes
Requirements: Required Qualifications
* Technical Background: M.S. or Ph.D. in data Science, Machine Learning, Computer Science, or equivalent practical experience in deep learning.
* RL & Post-Training Expertise: Strong hands-on experience training large-scale models using RL algorithms (e.g. GRPO, PPO) on open-weight architectures.
* AI Security Expertise: Solid understanding of LLM vulnerabilities, red-teaming methodologies, and defensive alignment against IPI attacks.
* Infrastructure Skills: Proficiency in PyTorch, Docker containerization, and distributed training architectures.
* Diagnostic Skills: Ability to analyze agent rollout traces, craft deterministic rubrics/verifiers, and debug complex reward shaping flaws. Preferred Qualifications
* Prior experience working with standard RL gym formats, such as Harbor.
* Experience evaluating complex agentic workflows in tool-use or web-browser environments.
* Familiarity with evaluating open-weight models similar to Llama or Mistral against adversarial workloads.
This position is open to all candidates.