← News

Agents & MCPSources: arXiv · 16 Nov 2026

LLM verifier judgments can flip when agent execution evidence crosses abstraction boundaries

A new arXiv paper introduces Verification Transport, a test of whether LLM-based verifier verdicts stay stable when the same agent execution is presented through tools, delegation boundaries or compacted histories.

In plain words

AI systems that judge whether an AI agent did a task correctly usually see the agent's actions in one fixed format. The authors show that reformatting the same actions — for example by summarising them or hiding steps behind delegation — can change the verdict. They propose a test, Verification Transport, that measures how often judgments survive such changes.

What was announced

The paper, an arXiv abstract by its authors, defines Verification Transport as a reliability test that measures judgment stability under abstraction-induced changes in execution observability. The method builds matched native, abstracted and restored views of the same frozen executions, holding the task, trajectory, outcome, ground truth, evaluation criterion and verifier fixed, so only the observability of execution evidence varies.

Across four agent-evaluation benchmarks, seven verifier backbones, four verification protocols and three abstraction boundaries, the authors report substantial transport failure, reaching about 31% under macro abstraction and delegation. Direct verification more often produced incorrect verdicts, while evidence-grounded and deliberative protocols tended to abstain under limited observability, which the authors say means lower failure rates do not necessarily imply greater reliability. Restoring the corresponding execution provenance recovered 85.5% of transport failures, with lower recovery when provenance was lossy or unavailable. The authors conclude that native-view accuracy alone is insufficient to characterise verifier reliability.

What it does not show

  • Evidence is limited to the paper's abstract; full experimental details, benchmarks and methods are not available in the supplied passages.
  • The reported figures are the authors' own measurements on their chosen benchmarks and verifiers; no independent confirmation is provided.

For security and IT teams

Scope
The findings concern LLM-based verifiers evaluating agent executions across abstraction boundaries such as composed tools, delegation and compacted histories, not verifier accuracy in a single fixed representation.

Sources

Source-support review completed 17 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 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.