Search papers, labs, and topics across Lattice.
2
2
5
6
Training LLMs on a massive, synthesized dataset of verified Rust programs slashes the cost of generating formal correctness proofs, outperforming even Claude Sonnet 4.5.
Train massive models on unproven hardware with confidence: SIGMA achieves 94% accelerator utilization and trains a 200B MoE with near-perfect stability on early-life AI accelerators.
