In this paper, we propose LeGEND that features a top-down style of scenario generation for testing of the ADS that starts with functional scenarios and then steps downwards. Specifically, LeGEND employs two LLMs to transform functional scenarios to formal logical scenarios, and then searches with the logical scenarios for critical scenarios. To avoid unrealistic scenarios at the functional level, LeGEND selects functional scenarios from real-world accident reports and employs an intermediate representation, called interactive pattern sequence, to record the featured events and their logical relations in reports. Experimental results on the industrial-grade ADS, Baidu Apollo, demonstrate that LeGEND can generate a more diverse set of critical scenarios compared to existing approaches.