Lead LLM‑Driven Adversarial Curriculum Engineer

corpora-jobs-1778796293285-db9d41c6 - Frontier Development

Applied ScientistLead1 position

⚡

Why This Role is Different

Frontier Development Role

Design and implement a self‑learning LLM system that crafts semantic adversarial prompts to stress‑test multi‑agent policies, driving robust learning and uncovering hidden failure modes.

The Frontier Element

You’ll create an LLM‑based red‑team that can autonomously generate instruction‑level attacks, a frontier concept in AI safety and policy robustness that pushes beyond gradient‑based adversaries.

🔬

Project Context

Research Area

LLM‑Driven Adversarial Curriculum (LLM‑AC) for semantic adversarial scenario generation

From: Adversarial Observation Perturbations and Policy Inference

Why This Role is Critical

Key to expose policy brittleness through high‑level semantic attacks and to generate diverse, maximally regretful scenarios.

What You Will Build

Attacker‑target‑judge loop harnessing LLM‑TOC, retrieval‑augmented generation pipeline, curriculum scheduling, and integration with MARL training.

🛠

Key Responsibilities

Build the attacker‑target‑judge loop where an LLM generates jailbreak or policy‑shifting prompts for target agents.
Integrate retrieval‑augmented generation (RAG) to ground semantic attacks in realistic knowledge bases.
Design curriculum scheduling that maximizes regret for MARL agents while ensuring coverage of unseen semantic perturbations.
Automate evaluation of policy brittleness and trigger curriculum updates in real time.
Collaborate with the Bayesian Policy Inference team to incorporate semantic attack signals into the policy prior.

🎯

Required Skills & Experience

Technical Must-Haves

Large Language Model fine‑tuning and prompt engineering

Expert

Core to generating high‑quality semantic adversarial scenarios.

Retrieval‑augmented generation (RAG) pipelines

Advanced

Ensures semantic attacks are grounded in accurate knowledge.

Multi‑agent reinforcement learning (MARL) and curriculum learning

Advanced

Needed to evaluate and adapt policy responses.

Python, PyTorch, HuggingFace Transformers

Expert

Daily stack for LLM development.

Adversarial AI and AI safety principles

Proficient

Guides responsible generation of harmful prompts.

Experience Requirements

5+ years in NLP or AI safety research.
Published work on LLM‑based adversarial generation or red‑team frameworks.
Experience building automated curriculum systems for RL.

Education

PhD or MS in NLP, AI Safety, or Robotics with experience in LLM fine‑tuning and adversarial generation.

⭐

Preferred Skills

Knowledge of LLM‑based jailbreak detection and mitigation.
Experience with multi‑modal LLMs (vision‑language) for semantic attack generation.
Background in human‑in‑the‑loop interpretability for policy debugging.

🤝

You Will Thrive Here If...

Comfortable designing systems that generate potentially harmful content for defensive purposes.
Eager to iterate rapidly on LLM prompts and curriculum schedules.
Passionate about pushing the frontiers of AI safety and robustness.

📈

Impact & Growth

12-Month Impact

Within a year, produce a curriculum that increases policy resilience by 40% against unseen semantic attacks in benchmark multi‑agent environments, directly enhancing mission reliability.

Growth Opportunity

Expand the curriculum framework to other domains (e.g., autonomous driving, cyber‑defense) and lead a growing team of LLM engineers focused on adversarial robustness.

Ready to Push the Boundaries?

If this sounds like the challenge you have been looking for, we want to hear from you. We value what you can build over where you have been.