We use an example, to illustrate how LeGEND transforms a real accident report to a logical scenario. The description of the accident as well as an illustration are shown in the left figure. The accident involves three vehicles, denoted as V1, V2 and V3, driving in three different lanes at the initial stage. Then, V1 wants to change to lane 1 by overtaking V2, but due to their close distance, V1 collides with V2; due to the impact the collision, V2 swerves left and hits V3 in lane 3 subsequently. 

In Phase 1, LLM1 processes the above accident report  and transforms it to an intermediate representation called interactive pattern sequence (IPS), as shown in the left figure. Each IPS consists of a tuple ⟨DI , DP⟩, where DI denotes the initial actions of vehicles, and DP describes the interactive patterns between vehicles before accident. 

• DI is a sequence initial actions, where Vi is a vehicle, followed by a natural language description about the states and the intended action of Vi . 

                            ⟨Vi⟩ : { initial states and actions of Vi } 

• DP  is a sequence of interactive patterns, where Vi and Vj are two vehicles, followed by a description about the actions of Vi and Vj w.r.t. their interaction: 

                  ⟨Vi ,Vj⟩ : {actions of Vi and Vj related to interaction} 

Additionally, the extracted interactive pattern sequences undergo a syntactic legality check to ensure correct formatting as stated above. This step primarily verifies that the elements in ⟨⟩ and ⟨,⟩ are all vehicles and that there are two vehicles participating in interactive patterns. This is necessary because the LLM might incorrectly extract interactive patterns, e.g., involving one vehicle and a static object such as a stone or road boundary. 

In Phase 2, LLM2 converts the extracted IPS to a logical scenario expressed by a formal DSL, as shown in the left figure. In this DSL, a test case S is a sequence of initial states of vehicles and actions of NPC vehicles, as follows: 

                                             S = { I1, . . . , Im , A1, . . . , An}

where Ii ( i ∈ {1, . . . , m} ) is the initial states of the i-th vehicle, and Aj (j ∈ {1, . . . , n} ) is an action of an NPC vehicle. During the template generation process, the names of the functions in S are fixed, while the parameters are unspecified. 

Next, the generated template is populated with parameter ranges by LLM2, using information about road specifications and default parameters. However, since the parameter ranges filled by LLMs may not be entirely rational (see RQ3), LeGEND uses a combination of LLM-supplied and default parameter ranges for some continuous variables (e.g., the start positions of vehicles): 

• We first check if the parameter ranges generated by the LLM are valid. This mainly involves checking the validity of initial positions of vehicles: if two vehicles are in the same lane, we require that the minimum possible distance between them exceeds a certain threshold to avoid an initial collision; if two vehicles are not in the same lane, there is no such constraint;

• If the ranges generated by the LLM pass this check, we adopt them because they are more precise and so more efficient for identifying critical scenarios;

• Otherwise, we use the default ranges, which are predefined values retrieved from existing map files (details available in our code repository).