TY - GEN
T1 - Syntax Is Easy, Semantics Is Hard
T2 - 2026 ACM Secure Development Conference, SecDev 2026
AU - Danso, Priscilla Kyei
AU - Hasan, Mohammad Saqib
AU - Balasubramanian, Niranjan
AU - Chowdhury, Omar
N1 - Publisher Copyright:
© 2026 Copyright held by the owner/author(s).
PY - 2026/6/30
Y1 - 2026/6/30
N2 - Propositional Linear Temporal Logic (LTL) is a popular formalism for specifying desirable requirements and security and privacy policies for software, networks, and systems. Yet expressing such requirements and policies in LTL remains challenging because of its intricate semantics. Since many security and privacy analysis tools require LTL formulas as input, this difficulty places them out of reach for many developers and analysts. Large Language Models (LLMs) could broaden access to such tools by translating natural language fragments into LTL formulas. This paper evaluates that premise by assessing how effectively several representative LLMs translate assertive English sentences into LTL formulas. Using both human-generated and synthetic ground-truth data, we evaluate effectiveness along syntactic and semantic dimensions. The results reveal three findings: (1) in line with prior findings, LLMs perform better on syntactic aspects of LTL than on semantic ones; (2) they generally benefit from more detailed prompts; and (3) reformulating the task as a Python code-completion problem substantially improves overall performance. We also discuss challenges in conducting a fair evaluation on this task and conclude with recommendations for future work.
AB - Propositional Linear Temporal Logic (LTL) is a popular formalism for specifying desirable requirements and security and privacy policies for software, networks, and systems. Yet expressing such requirements and policies in LTL remains challenging because of its intricate semantics. Since many security and privacy analysis tools require LTL formulas as input, this difficulty places them out of reach for many developers and analysts. Large Language Models (LLMs) could broaden access to such tools by translating natural language fragments into LTL formulas. This paper evaluates that premise by assessing how effectively several representative LLMs translate assertive English sentences into LTL formulas. Using both human-generated and synthetic ground-truth data, we evaluate effectiveness along syntactic and semantic dimensions. The results reveal three findings: (1) in line with prior findings, LLMs perform better on syntactic aspects of LTL than on semantic ones; (2) they generally benefit from more detailed prompts; and (3) reformulating the task as a Python code-completion problem substantially improves overall performance. We also discuss challenges in conducting a fair evaluation on this task and conclude with recommendations for future work.
KW - Formal methods
KW - LLMs
KW - NL-to-LTL
KW - Security/Privacy policy
UR - https://www.scopus.com/pages/publications/105045273533
U2 - 10.1145/3805773.3806005
DO - 10.1145/3805773.3806005
M3 - Conference contribution
AN - SCOPUS:105045273533
T3 - SecDev 2026 - Proceedings of the 2026 ACM Secure Development Conference, Co-located with FSE 2026
SP - 152
EP - 168
BT - SecDev 2026 - Proceedings of the 2026 ACM Secure Development Conference, Co-located with FSE 2026
A2 - Kannavara, Raghudeep
A2 - Celik, Z. Berkay
A2 - Rahaman, Sazzadur
A2 - Van Landuyt, Dimitri
PB - Association for Computing Machinery, Inc
Y2 - 5 July 2026 through 6 July 2026
ER -