Sreehari Sankar and co-authors submitted LeanGuard to arXiv on 2 November 2026. The preprint describes a runtime that checks each proposed tool call against a policy written in Lean4 before forwarding it to the environment. The authors report eliminating calls forbidden by their Lean policy in a Tau-Bench evaluation.[1]

According to the abstract, policies combine conditions on state with scoped temporal operators that refer to execution history. Evaluation is fail-closed, and the authors say their decision procedure is mathematically proved equivalent to a LeanLTL semantics. After a denial, feedback can identify the violated rule and offer a hint. A separate repair mechanism uses Lean metaprogramming to search for a nearby admissible call and prove policy admission without using the model.[1]

On Tau-Bench with five Qwen models ranging from 0.8 billion to 27 billion parameters, the authors report at least 33 forbidden calls from unguarded agents and none from LeanGuard; task success with feedback did not change significantly. On DTap, they report direct prompt-injection success falling from 57.9% to 5.9%. These are results reported in the abstract, whose stated policy and evaluation scope matter when interpreting them.[1]

[1]: LEANGUARD: Lean-Verified Guardrails for Tool-Using Agents, arXiv, 2 November 2026.

Sources

  1. LEANGUARD: Lean-Verified Guardrails for Tool-Using AgentsarXiv · 2 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

8cb0c16199b309e730cc0b934b09684fe32fd2c39293954ed40c941174f278e2