arxivcs.CRcs.LG2026-06-30
Certified Speculative Execution for Untrusted AI Agents
Chenyu Zhou, Qiliang Jiang, Shuning Wu, Xu Zhou
Hard-constrained sequential decision systems have no certified way to spend the test-time compute of modern AI: executing the multi-step drafts of a learned policy or a frozen LLM forfeits the feasibility guarantee a trusted solver provides, while invoking the solver at every ste…