Work with us
Environments in our post-training and evaluation tasksets are built to push realistic complexity further than what is typically used.
We use techniques like multi-agent to simulate many humans and/or collaborators, dozens of cloned services with realistic messiness, and human expert QA for reward hacking and environment robustness.
Pre-Deployment Model Evaluations
- Multi-agent social behavior and capability evaluations
- Anonymized checkpoints on Multi-Agent Arena
- Simulated real-world scenarios
Real-World Agentic Capability
- Engineering inside realistic enterprises with many services
- Game development
- Robot-use simulations
Alignment and Safety
- Multi-agent social games
- Real-world enterprise scenarios
- Red-teaming and adversarial training
Get in touch
Talk to our team about your areas of focus for training and evaluations. We are familiar with the communication, standards, and packaging that research teams working with us expect.