Transfer across workflow-author models

Investigate whether PolicyGuide's performance transfers across workflows authored by different language models, rather than relying on a single GPT 5.4-authored workflow graph.

Background

The experiments reuse one GPT 5.4-authored Airline workflow graph unchanged across GPT 5.4, Claude Sonnet 4.6, and Gemini 2.5 Pro agents, thereby testing executor-side transfer while holding the workflow representation fixed. This design does not test whether the method remains effective when the workflow itself is generated by a different model or authoring process. Establishing such transfer would address the generality of PolicyGuide's offline workflow-generation component, which is distinct from transfer across agent families.

References

Transfer across workflow-author models remains untested.

— PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents  (2608.19861 - Kang et al., 20 Aug 2026) in Section 5, subsection “Generalization across agent families”

Several directions remain open for future study. (i) We use a fixed external model as a scalable proxy for human workflow designers, and future work should evaluate guidance written by people with more diverse intentions and error patterns.

— Thinking Outside the Box: Can Language Models Rely on External Guidance Selectively?  (2609.39578 - Wang et al., 30 Sep 2026) in Conclusion