Research
The systems, methods, and experiments that shape how we build and safeguard AI products.
KillBench: discovering hidden biases of LLMs
Evaluating decision-making biases across frontier AI models.
CircleGuardBench: Evaluating AI Moderation Models
Measuring the real-world performance of AI guardrails.
NeurIPS 2026Coming soon
RISE — red-teaming text-to-image models at production scale
Coming soonPreprint · OptimizationComing soon
Hat-Muon: Communication-efficient distributed training methods
Coming soonInfrastructureComing soon
Custom kernels for Blackwell B300
Coming soonTrainingComing soon