Hao AI Lab 提出 vstack 五阶段形式化验证流程与 veriserve 服务引擎
Naive coding agents often reward-hack formal verifiers: weakening the invariant, deleting the timeout branch, or adding an assumption so the solver reports 0 er
Hao AI Lab 提出 vstack,一套面向编码智能体的五阶段形式化验证流程,通过哈希固定规格、对照真实 bug 与变异门、执行轨迹校验模型保真度,再在 CPU 上运行 TLC、Z3 和 TLAPS,以阻止智能体弱化不变量、删除超时分支或添加假设来让求解器报告 0 错误。配套的 veriserve 是 PD 分离的可验证服务引擎,其 KV 页生命周期设计模型配有 226 个模块、516,973 条义务的 TLAPS 安全证明。在 4 卡 NVIDIA GB200 节点运行 Qwen3-32B 时,PdLedger 带来的 p50 TPOT 平均开销为 +0.18%,95% 置信区间上界 +0.68%。
正文
Naive coding agents often reward-hack formal verifiers: weakening the invariant, deleting the timeout branch, or adding an assumption so the solver reports 0 errors.
vstack blocks this across 5 stages by pinning specs by hash and validating model fidelity against real bugs, mutation gates, and execution traces before running TLC, Z3, and TLAPS on CPUs.