Automated Repair of Ambiguous Natural Language Requirements

The rise of large language models (LLMs) has amplified the role of natural language (NL) in software engineering, and its inherent ambiguity and susceptibility to misinterpretation pose a fundamental challenge for software quality, because employing ambiguous requirements may result in the generation of faulty programs. The complexity of ambiguity detection and resolution motivates us to introduce the problem of automated repair of ambiguous NL requirements. Repairing ambiguity in requirements poses a challenge for LLMs, as it demands a metacognitive capability - the ability to reflect on how alterations to the text influence their own interpretation of this text. Indeed, our experiments show that directly prompting an LLM to detect and resolve ambiguities results in irrelevant or inconsistent clarifications. Our key novelty is in decomposing this problem into simpler subproblems which do not require metacognitive reasoning. First, we analyze and repair LLM's interpretation of requirements embodied in the distribution of programs they induce using traditional testing and program repair methods. Second, we repair requirements based on the changes to the distribution via what we refer to as contractive specification inference. This decomposition enables targeted, minimal requirement repairs that yield cross-model performance gains in code generation. We implemented this approach in a tool SpecFix, and evaluated it using three SOTA LLMs, GPT-4o, DeepSeek-V3 and Qwen2.5-Coder-32b-Instruct, across two widely-used code generation benchmarks: HumanEval+ and MBPP+. Our results show that SpecFix, operating autonomously without human intervention or external information, outputs repaired requirements that, when used by LLMs for code generation, increase the Pass@1 score by 4.3%, and help LLMs to solve 3.4% more problems via majority vote.