Verification of Faithfulness in Automated Idea Execution Agents

Develop rigorous methods to verify the faithfulness and correctness of code implementations produced by an LLM-based idea execution agent, ensuring that baseline and proposed methods are implemented as specified and that evaluation metrics are computed correctly, rather than relying solely on final experiment outcomes.

Background

When attempting to automate execution of AI- or human-generated ideas, the authors find that agents often skip or alter baseline/proposed steps and sometimes implement incorrect metrics, producing misleading results even when the code runs.

They state that thorough implementation verification is necessary but non-trivial, and explicitly leave this verification problem to future work, highlighting the need for principled approaches to check faithfulness of agent-produced code.