RefinementGuard validates hidden agent-tool execution graphs before runtime admission
The arXiv paper introduces translation validation for agent operations lowered into hidden parallel graphs, paired with fail-closed admission that rejects unqualified runs.
In plain words
AI agents call tools, and the runtime may secretly rewrite those calls into a parallel execution graph that is faster but can change what the operation actually does. RefinementGuard checks the hidden graph against an independent reference before letting it run, and blocks it if the check fails. In the authors' tests it approved all valid variants while catching about half of the injected faults.
What was announced
RefinementGuard, described in a new arXiv paper, targets a specific risk: tool runtimes that lower an agent-selected operation into a hidden parallel graph, which reduces latency and context but risks semantic mistranslation. The system performs translation validation on these hidden, effectful lowerings and couples it to artifact-bound, fail-closed runtime admission. Qualification compares a candidate graph with an independent reference on isolated state, using explicit parent-operation obligations.
The authors evaluate two repository-level contracts, bug localization and patch validation, on controlled repositories. Among 100 activated, non-equivalent faults adjudicated by one human, the validation rejected 49, with a reported 95% confidence interval of 33.7% to 64.8%; the stated goal is selective admission rather than maximum rejection, since exact equality would also reject all 60 valid variants while obligation-based validation accepted all 60. On a frozen held-out split, results are localization-only because 21 patch mutants failed structurally, with the B3 configuration rejecting 38 of 56 localization defects. In a controlled asynchronous-delay regime, the localization median-latency ratio between serial and guarded execution was 4.50x, and hidden execution exposed 74.3% less parent-visible context. The paper reports no demonstrated speedup for patch validation, where zero-delay runs showed overhead.
What it does not show
- Evidence is limited to the paper's abstract; the evaluation uses controlled repositories, one human adjudicator, and a controlled asynchronous-delay regime, and the held-out patch-validation result was incomplete because 21 patch mutants failed structurally.
For security and IT teams
- Scope
- The reported evaluation covers two repository-level contracts, bug localization and patch validation, on controlled repositories rather than production workloads.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.