Researcher, CoT Monitorability & Alignment

Society for AI Security; San Francisco (CA); Hybride; USD 210 000 - 270 000; Plein temps

Avantages offerts par ce poste

Relocation assistance

Hybrid work model (3 days in office)

Résumé du poste

OpenAI is seeking a researcher on the Alignment team to design and run experiments that improve understanding of model monitorability. You will investigate how training interventions influence whether reasoning remains legible and help translate findings into practical oversight.

You will collaborate with researchers and engineers across model training, alignment evaluations, monitoring, and frontier-risk work, producing externally publishable research when advances occur.

Qualifications

• Strong hands-on experience training, evaluating, or debugging large ML models, especially LLMs.
• Deep curiosity and interest in alignment, interpretability, or model behavior.

Responsabilités

• Design and run experiments to understand model monitorability across frontier reasoning models and training settings.
• Build evaluations to measure whether monitors predict properties of interest, including high-stakes misbehavior.
• Investigate interventions in pre-training, mid-training, post-training, RL and others to affect monitorability.
• Analyze model behavior and translate observations into hypotheses, experiments, and recommendations.
• Translate findings into practical monitoring and oversight approaches for real training runs.
• Collaborate with researchers and engineers across model training, alignment evaluations, monitoring, and frontier-risk work.
• Produce externally publishable research when results advance alignment science.

Connaissances

Empirical ML

Model behavior

Interpretability

LLMs

Experiment design

Description du poste

OpenAI is seeking a researcher on the Alignment team to design and run experiments that improve understanding of model monitorability. You will investigate how training interventions influence whether reasoning remains legible and help translate findings into practical oversight.

You will collaborate with researchers and engineers across model training, alignment evaluations, monitoring, and frontier-risk work, producing externally publishable research when advances occur.