Automated LTL Specification Generation from Industrial Aerospace Requirements

Zhi Ma, Xiao Liang, Cheng Wen, Rui Chen, Bin Gu, Shengchao Qin, Cong Tian, Mengfei Yang · 2026

Read Automated LTL Specification Generation from Industrial Aerospace Requirements by Zhi Ma, Xiao Liang, Cheng Wen, Rui Chen, Bin Gu, Shengchao Qin, Cong Tian, Mengfei Yang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-21715.txt.

Introduction

Ma and seven colleagues automate a slow, expert-heavy job in aerospace software: turning plain-language requirements into Linear Temporal Logic properties for safety checks. Their AeroReq2LTL system first normalises jargon through a data dictionary. It then makes time cues and logic links explicit in a template language before translation. On a real aerospace dataset, it reaches 85 percent precision and 88 percent recall, with outputs ready for existing checkers.

Catalog record

CREATORZhi Ma, Xiao Liang, Cheng Wen, Rui Chen, Bin Gu, Shengchao Qin, Cong Tian, Mengfei Yang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47192
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.21715

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Automated LTL Specification Generation from Industrial Aerospace Requirements Zhi Ma, Xiao Liang, Cheng Wen, Rui Chen, Bin Gu, Shengchao Qin, Cong Tian, Mengfei Yang Abstract In the development and verification of safety-critical aero-space software, Linear Temporal Logic (LTL) has been widely used to specify complex system properties derived from requirements. However, a significant gap remains in industrial practice: translating natural language (NL) requirements into formal LTL properties is a labor-intensive and error-prone process that requires rare expertise in both aerospace control engineering and formal methods. While recent NL-to-LTL tools ( e.g. , NL2SPEC, NL2TL, NL2LTL) are capable of automating parts of this process, they often fail on real requirement documents in industrial settings, due to complex domain terminology or implicit temporal and logical structure. To address these challenges, we present AeroReq2LTL , a framework that automates LTL property generation for aerospace requirements using large language models (LLMs), with two key industrial innovations: (i) a data dictionary that normalizes technical jargon into precise atomic propositions; and (ii) a template-based requirement language that makes temporal cues and logical relations explicit before translation. On a real aerospace dataset, AeroReq2LTL achieves 85% precision and 88% recall in LTL generation, and its outputs can be directly consumed by existing verification tools. 1 Introduction Software verification for safety-critical aerospace control software must provide strong and auditable assurance under tight engineering constraints, including frequent requirement changes, limited verification expertise, and strict expectations on evidence quality. In such settings, formal methods are often adopted because they can offer mathematically grounded guarantees beyond what traditional testing can typically provide. Within this domain, Linear Temporal Logic (LTL) [ 30 ] has emerged as the preferred formal specification language for expressing complex system properties. Because aerospace software is inherently reactive—interacting continuously with an environment over time—its correctness depends not only on functional behavior but also on precise temporal properties [ 25 ] . LTL provides the mathematical rigor necessary to express critical safety invariants and liveness conditions [ 2 ] , which are widely supported by existing model checkers and runtime verification tools. Despite these advantages, the industrial adoption of formal methods is impeded by the formalization bottleneck: the manual translation of natural-language (NL) requirements into precise LTL properties. This process is labor-intensive and error-prone, requiring rare expertise in both aerospace control engineering and formal logic. Furthermore, recent research exploring Large Language Models (LLMs) [ 26 , 27 , 35 , 32 ] for this task ( e.g. , NL2SPEC [ 11 ] , NL2TL [ 10 ] , NL2LTL [ 14 ] ) has highlighted a fundamental limitation: while these tools succeed on synthetic benchmarks, they often fail on real industrial documents (as shown in our motivating example and experiments under the industrial setting). The core challenge lies in the heterogeneous nature of industrial requirements. In practice, aerospace requirement documents contain highly specialized jargon, ambiguous abbreviations [ 3 ] , and implicit temporal and logical structures that are not explicitly stated in the text. Consequently, existing tools often produce incorrect or incomplete specifications, limiting their applicability in practice. To address these challenges, we present AeroReq2LTL , a systematic engineering framework designed to automate the generation of LTL specifications from industrial aerospace requirements. AeroReq2LTL leverages the reasoning capabilities of LLMs but enhances them through two key industrial innovations: • SpaceKG : A domain-specific data dictionary that automatically ingests engineering artifacts (such as interface tables) to resolve technical jargon into precise, code-level atomic propositions. • SpaceRDL : A structured template language that guides LLMs to uncover implicit temporal keywords and logical relations within the context of aerospace control logic. Designed for seamless toolchain integration, AeroReq2LTL facilitates a direct connection with the subsequent verification tool to support batch processing and iterative feedback loops. The framework was evaluated using 79 production requirements from real-world spacecraft control systems. Experimental results demonstrate that AeroReq2LTL achieves 85% precision and 88% recall, substantially reducing manual formalization effort and establishing a practical, automated pathway for rigorous software assurance within the aerospace industry. 2 Background and Motivation 2.1 The complexity of Industrial Requirements Table 1 : Examples of Natural Language and Corresponding LTL Specifications Natural Language LTL Specification [ 10 ] Once red, the light cannot become green next. G(\text{red}\rightarrow X\neg\text{green}) [ 10 ] Once the light is red, it must remain red until it turns yellow. G(\text{red}\rightarrow\text{red}\,U\,\text{yellow}) [ 11 ] If a holds then c is true until b. \text{a}\rightarrow(\text{c}\,U\,\text{b}) [ 11 ] If b holds, next c holds until a holds or always c holds. \text{b}\rightarrow X((\text{c}\,U\,\text{a})\lor G\,\text{c}) [ 14 ] Navigate to the green room while avoiding landmark 1. (F\,\text{green})\land G(\neg\text{landmark~1}) [ 14 ] Swing by landmark 1 before ending up in the red room. F(\text{landmark~1}\land(F\,\text{red})) In academic research, the task of translating natural language to Linear Temporal Logic (LTL) is often simplified to mapping short, isolated sentences. Table 1 presents six representative examples adopted in prior work, including NL2LTL [ 14 ] , NL2SPEC [ 11 ] , and NL2TL [ 10 ] . In these examples, the correspondence between natural language statements and LTL specifications is explicit and largely unambiguous, rendering the translation task relatively straightforward. However, industrial requirements are fundamentally different, as illustrated by the example simplified from the ACS-LEOS 1 1 1 The Attitude Control Software for Low Earth Orbit Satellites (ACS-LEOS) is responsible for processing sensor data from sunlight and infrared instruments, managing transitions among multiple operating modes, and executing attitude adjustment tasks such as three-axis stabilization, spin stabilization, and sun-pointing control. in Fig. 1 . They are not merely free-form prose but heterogeneous engineering artifacts characterized by three distinct layers of complexity: (1) Multiple types of context : A single functional requirement ( e.g. , Req13 ) cannot be parsed in isolation. Its meaning is bound to the Interface Description , which specifies port-level information such as signal names, data types, initial values, and valid ranges, as well as the Invocation Condition , which specifies the 128 ms control cycle. (2) Domain-Specific Semantics: Technical terminologies often act as placeholders for complex logic. In Req13 , the phrase “absolute values of the angular rates on all axes are less than 0.15°/s” is an empirical condition for spacecraft stability. Without the domain knowledge found in the interface port explanations, an automated tool cannot map this intent to the correct variable, dwCount . (3) Implicit Temporal Logic: Industrial requirements often omit explicit temporal operators, for example, in Req13 , the term "switches from…to" implies a state transition that must occur at the next time step ( X ) once a sustaining condition is met. ((a)) The requirement description for the Mode Switch Function Module of ACS-LEOS. ((b)) Misalignment of LLM-Generated LTL with Industrial Requirements. Figure 1 : The complexity and accuracy in NL requirement and its specification. 2.2 Limitations of General-Purpose LLMs While LLMs like GPT-4 possess powerful reasoning capabilities, the generated LTL specifications are often unsatisfactory [ 9 , 36 ] . As shown in Fig. 1 , although the model attempts to incorporate all relevant information in the requirement ( Req13 ), the generated LTL specification deviates substantially from the intended semantics. This mismatch stems primarily from three primary reasons. (1) Contextual Blindness to Engineering Artifacts. LLMs struggle to connect natural-language requirements

Read the complete text