Research Scientist Graduate (Seed Model-Alignment&Preparedness) - 2027 Start

ByteDance, San Jose, California

What you will do

Agent RL & safety training: Use RL, SFT, and preference optimization to train agents safely. Design environments, rewards, and training data that reduce deception, reward hacking, and unauthorized actions, and teach agents to check authorization, recognize risks, and stop when needed, without sacrificing useful autonomy. Ensure safety generalizes to new tasks and holds up through further training.
Frontier risk assessment: Partner with domain experts to evaluate dangerous capabilities (e.g., CBRN, cybersecurity), model threats, and build safeguards. Provide empirical evidence to guide development and deployment decisions.
Automated red teaming: Build automated attack generation and adversarial training methods to harden models against jailbreaks, prompt injection, and other adversarial inputs.
Monitoring & oversight: Detect risks from inputs and outputs, execution traces, chains of thought, and internal representations. Build oversight models, action review, and permission controls, and measure their effectiveness, evasion resistance, and cost.

ByteDance, San Jose, California

Agent RL & safety training: Use RL, SFT, and preference optimization to train agents safely. Design environments, rewards, and training data that reduce deception, reward hacking, and unauthorized actions, and teach agents to check authorization, recognize risks, and stop when needed, without sacrificing useful autonomy. Ensure safety generalizes to new tasks and holds up through further training.
Frontier risk assessment: Partner with domain experts to evaluate dangerous capabilities (e.g., CBRN, cybersecurity), model threats, and build safeguards. Provide empirical evidence to guide development and deployment decisions.
Automated red teaming: Build automated attack generation and adversarial training methods to harden models against jailbreaks, prompt injection, and other adversarial inputs.
Monitoring & oversight: Detect risks from inputs and outputs, execution traces, chains of thought, and internal representations. Build oversight models, action review, and permission controls, and measure their effectiveness, evasion resistance, and cost.