Automaton-Based Representations of Task Knowledge from Generative Language Models

Yunhao Yang, Cyrus Neary, Ufuk Topcu
Proceedings of the International Conference on Neuro-symbolic Systems, PMLR 288:765-783, 2025.

Abstract

Automaton-based representations play an important role in control and planning for sequential decision-making problems. However, obtaining the high-level task knowledge required to build such automata is often difficult. Meanwhile, generated language models (GLMs) can automatically generate relevant text-based task knowledge. However, such text-based knowledge cannot be formally verified or used for sequential decision-making. We propose a novel algorithm named GLM2FSA that constructs a finite state automaton (FSA) encoding high-level task knowledge from a brief natural-language description of the task goal. GLM2FSA first sends queries to a GLM to extract task knowledge in textual form, and then it builds an FSA to represent this text-based knowledge. The proposed algorithm thus fills the gap between natural-language task descriptions and automaton-based representations, and the constructed FSAs can be formally verified against user-defined specifications. We accordingly propose a procedure to iteratively refine the input queries to the GLM based on the outcomes, e.g., counter-examples, from verification. We apply the proposed algorithm to an autonomous driving system to demonstrate its capability for sequential decision-making and formal verification. Furthermore, quantitative results indicate the refinement method improves the probability of generated knowledge satisfying the specifications by 40 percent.

Cite this Paper

BibTeX
@InProceedings{pmlr-v288-yang25a, title = {Automaton-Based Representations of Task Knowledge from Generative Language Models}, author = {Yang, Yunhao and Neary, Cyrus and Topcu, Ufuk}, booktitle = {Proceedings of the International Conference on Neuro-symbolic Systems}, pages = {765--783}, year = {2025}, editor = {Pappas, George and Ravikumar, Pradeep and Seshia, Sanjit A.}, volume = {288}, series = {Proceedings of Machine Learning Research}, month = {28--30 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v288/main/assets/yang25a/yang25a.pdf}, url = {https://proceedings.mlr.press/v288/yang25a.html}, abstract = {Automaton-based representations play an important role in control and planning for sequential decision-making problems. However, obtaining the high-level task knowledge required to build such automata is often difficult. Meanwhile, generated language models (GLMs) can automatically generate relevant text-based task knowledge. However, such text-based knowledge cannot be formally verified or used for sequential decision-making. We propose a novel algorithm named GLM2FSA that constructs a finite state automaton (FSA) encoding high-level task knowledge from a brief natural-language description of the task goal. GLM2FSA first sends queries to a GLM to extract task knowledge in textual form, and then it builds an FSA to represent this text-based knowledge. The proposed algorithm thus fills the gap between natural-language task descriptions and automaton-based representations, and the constructed FSAs can be formally verified against user-defined specifications. We accordingly propose a procedure to iteratively refine the input queries to the GLM based on the outcomes, e.g., counter-examples, from verification. We apply the proposed algorithm to an autonomous driving system to demonstrate its capability for sequential decision-making and formal verification. Furthermore, quantitative results indicate the refinement method improves the probability of generated knowledge satisfying the specifications by 40 percent.} }
Endnote
%0 Conference Paper %T Automaton-Based Representations of Task Knowledge from Generative Language Models %A Yunhao Yang %A Cyrus Neary %A Ufuk Topcu %B Proceedings of the International Conference on Neuro-symbolic Systems %C Proceedings of Machine Learning Research %D 2025 %E George Pappas %E Pradeep Ravikumar %E Sanjit A. Seshia %F pmlr-v288-yang25a %I PMLR %P 765--783 %U https://proceedings.mlr.press/v288/yang25a.html %V 288 %X Automaton-based representations play an important role in control and planning for sequential decision-making problems. However, obtaining the high-level task knowledge required to build such automata is often difficult. Meanwhile, generated language models (GLMs) can automatically generate relevant text-based task knowledge. However, such text-based knowledge cannot be formally verified or used for sequential decision-making. We propose a novel algorithm named GLM2FSA that constructs a finite state automaton (FSA) encoding high-level task knowledge from a brief natural-language description of the task goal. GLM2FSA first sends queries to a GLM to extract task knowledge in textual form, and then it builds an FSA to represent this text-based knowledge. The proposed algorithm thus fills the gap between natural-language task descriptions and automaton-based representations, and the constructed FSAs can be formally verified against user-defined specifications. We accordingly propose a procedure to iteratively refine the input queries to the GLM based on the outcomes, e.g., counter-examples, from verification. We apply the proposed algorithm to an autonomous driving system to demonstrate its capability for sequential decision-making and formal verification. Furthermore, quantitative results indicate the refinement method improves the probability of generated knowledge satisfying the specifications by 40 percent.
APA
Yang, Y., Neary, C. & Topcu, U.. (2025). Automaton-Based Representations of Task Knowledge from Generative Language Models. Proceedings of the International Conference on Neuro-symbolic Systems, in Proceedings of Machine Learning Research 288:765-783 Available from https://proceedings.mlr.press/v288/yang25a.html.

Related Material