Hao AI Lab releases verifiable LLM engine veriserve

Researchers at Hao AI Lab have introduced vstack, a formal verification workflow, and veriserve, a prototype verifiable LLM serving engine designed to catch silent RDMA and cache concurrency bugs in distributed inference. By modeling the system as a state machine and using TLA+ and TLAPS safety proofs, veriserve successfully eliminates data corruption risks during Prefill-Decode disaggregation with minimal TPOT overhead on NVIDIA GB200. Open article →