Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

A DbC Inspired Neurosymbolic Layer for Trustworthy Agent Design

Created by
  • Haebom

Author

Claudiu Leoveanu-Condrei

Outline

To address the lack of verifiability of the output of large-scale language models (LLMs), this paper proposes a contract layer that applies design-by-contract (DbC) and type-theoretic principles. This contract layer mediates all LLM invocations, specifies semantic and type requirements for inputs and outputs, and provides probabilistic corrections for compliance. This presents the LLM as a dual-viewpoint: a semantic parser and a probabilistic black-box component. Contract satisfaction is probabilistic, and semantic verification is operationally defined through programmer-specified conditions on well-defined data structures. Furthermore, this paper argues that two agents that satisfy the same contract are functionally equivalent for that contract.

Takeaways, Limitations

•
Takeaways:
◦
Improving LLM output reliability: The contract layer can improve the semantic correctness and type safety of LLM output.
◦
Defining functional equivalence of LLMs: A new perspective is presented, suggesting that LLMs that satisfy the same contract can be considered functionally equivalent.
◦
Clarifying the dual nature of LLM: Providing a new framework for understanding LLM as a semantic parser and a probabilistic black box.
◦
Programmable semantic verification: Semantic verification can be performed through conditions specified by the programmer.
•
Limitations:
◦
The Challenges of Contract Design: Designing a proper contract can require significant effort and expertise.
◦
Limitations of probabilistic satisfaction: Because contract satisfaction is probabilistic, perfect reliability cannot be guaranteed.
◦
Subjectivity of semantic verification: Semantic verification results may vary depending on conditions specified by the programmer.
◦
Performance overhead of the contract layer: potential performance degradation due to the addition of a contract layer.
👍
Made with Slashpage