Advancing the Foundations of Artificial Intelligence
Our research groups investigate verifiable test-time compute, ultra-long-context architectures, mechanistic interpretability, and alignment for frontier systems.
Scaling Verifiable Test-Time Compute in Alchemax-o3 Pro
How combining reinforcement learning with formal verification sandboxes enables monotonic capability scaling across mathematics, competitive programming, and systems engineering.
Alchemax-5 System Card & Frontier Safety Evaluation
Comprehensive evaluation of Alchemax-5 across cybersecurity, biological safety, autonomous replication, and adversarial robustness prior to general availability.
Exact Retrieval Across 2 Million Tokens with Ring-Sparse Attention
Eliminating context rot in ultra-long sequences through hardware-aware ring-sparse attention and exact key-value compression.
Mechanistic Interpretability of Multi-Step Agentic Tool Use
Mapping the internal sparse autoencoder features that govern goal persistence, error recovery, and uncertainty calibration in autonomous agents.