AI Trends

Trending story

Hao AI Lab launches vstack and veriserve

5 reports5 story updatesUpdated Nov 12 03:26

Understand this story

AI overview

On November 12, 2026, the UCSD Hao AI Lab released vstack, a five-stage formal verification workflow, and veriserve, a verifiable PD-disaggregated inference engine. vstack targets coding agents: it pins specifications via hashing, checks against real bugs and mutation gates, and validates model fidelity through execution traces, then runs TLC, Z3, and TLAPS on CPU to prevent agents from weakening invariants, deleting timeout branches, or adding assumptions so that the solver reports 0 errors. veriserve's KV page lifecycle design model comes with a TLAPS safety proof comprising 226 modules and 516,973 obligations; it uncovered 3 classes of real RDMA and cache races in the production engine and proposed 6 upstream fixes. When running Qwen3-32B on a 4-GPU NVIDIA GB200 node, PdLedger incurs an average p50 TPOT overhead of +0.18%, with a 95% confidence interval upper bound of +0.68%.

AI-generated from public reports

Latest development · Report

Hao AI Lab launches vstack and veriserve, providing formal verification for coding agents and inference services

Developments

1 confirmed developments

  1. Hao AI Lab launches vstack and veriserve

    Source:UCSD Hao AI Lab

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 12

  1. UCSD Hao AI LabFirst-party

    Hao AI Lab launches vstack and veriserve, providing formal verification for coding agents and inference services

    Hao AI Lab releases vstack and veriserve: the former is a 5-stage formal verification workflow for coding agents, and the latter is a verifiable PD-disaggregated inference serving engine. The team used vstack to discover 3 classes of RDMA and cache lifecycle defects in a production inference engine and proposed 6 upstream fixes; veriserve's KV page lifecycle design model comes with a TLAPS safety proof comprising 226 modules and 516,973 obligations. When running Qwen3-32B on a 4-GPU NVIDIA GB200 node, the average p50 TPOT overhead introduced by PdLedger is +0.18%, with a 95% CI upper bound of +0.68%.

  2. UCSD Hao AI LabFirst-party

    Hao AI Lab launches veriserve, a verifiable inference engine

    UCSD Hao AI Lab releases veriserve, a verifiable PD-disaggregated inference engine, which uses PdLedger to track each physical page with versioned (page, gen) tickets; pages that time out without evidence of stopping remain in QUARANTINED isolation, avoiding contamination of the KV cache. On 4-GPU NVIDIA GB200 with Qwen3-32B, PdLedger adds only 0.18% mean p50 TPOT overhead (95% CI upper bound +0.68%), intercepts all 3 injected glue faults, and replays 100% of 22,104 boundary events via TLC. The accompanying vstack five-stage formal verification process also found 3 classes of RDMA and cache races in the production engine and proposed 6 upstream fixes.

  3. UCSD Hao AI LabFirst-party

    Hao AI Lab launches vstack and veriserve verified inference engine

    Hao AI Lab has released vstack, a 5-stage formal verification pipeline for coding agents, and veriserve, a verifiable PD-disaggregated inference engine. They found 3 classes of RDMA and cache race conditions in production engines and proposed 6 upstream fixes. veriserve's KV page lifecycle design is backed by a TLAPS safety proof comprising 226 modules and 516,973 obligations, and when running Qwen3-32B on 4-GPU GB200, the average p50 TPOT overhead is only +0.18%.

  4. UCSD Hao AI LabFirst-party

    UCSD Hao AI Lab launches vstack and veriserve verifiable inference engine

    UCSD Hao AI Lab releases vstack, a five-stage formal verification workflow, and veriserve, a verifiable PD-disaggregated inference engine. It discovered 3 classes of real RDMA and cache races in production engines and proposed 6 upstream fixes. veriserve's KV-page lifecycle design model is backed by a TLAPS safety proof with 226 modules and 516,973 obligations, and when running Qwen3-32B on a 4-GPU GB200, the average p50 TPOT overhead is only +0.18%.

  5. UCSD Hao AI LabFirst-partySelected

    Hao AI Lab proposes the vstack five-stage formal verification pipeline and the veriserve serving engine

    Hao AI Lab proposes vstack, a five-stage formal verification pipeline for coding agents. It pins specifications via hashing, checks against real bugs and mutation gates, and validates model fidelity through execution traces; it then runs TLC, Z3, and TLAPS on CPU to prevent agents from weakening invariants, deleting timeout branches, or adding assumptions so that the solver reports 0 errors. The accompanying veriserve is a PD-disaggregated verifiable serving engine, whose KV page lifecycle design model comes with a TLAPS safety proof comprising 226 modules and 516,973 obligations. When running Qwen3-32B on a 4-GPU NVIDIA GB200 node, PdLedger incurs an average p50 TPOT overhead of +0.18%, with a 95% confidence interval upper bound of +0.68%.

Related stories

  1. Related storiesHao AI Lab open-sources the JetSpec++ model, code, and inference engine
  2. Related storiesEnglish translation is being prepared.
  3. Related storiesApple research team submits paper to arXiv disclosing the HALO system
  4. Related storiesDeepSeek open-sources inference framework DeepInfer