Do Large Language Models Correct False Technical Premises? An Empirical Case Study in Computer Science Question Answering

Abstract

Large language models (LLMs) are increasingly used for coding assistance, education, and technical question answering. A particular reliability problem arises when a question contains a false technical premise: instead of correcting the assumption, a model may accept it and construct a plausible answer around it. This study examines that behavior using a 60-prompt computer-science benchmark containing 30 factual questions and 30 false-premise questions across data structures, algorithms, operating systems, databases, networking, hardware, runtimes, and theoretical computer science. Each prompt was evaluated on Llama-3.1-8B-Instruct and Llama-3.3-70B-Versatile through the Groq inference API using temperature 0.0 and top-p 1.0, producing 120 responses. All responses were evaluated by the primary researcher using a predefined three-level rubric. Score 0 represents a correct answer or correct rejection of a false premise; Score 1 represents a substantially correct response with a limited factual problem; and Score 2 represents acceptance of the false premise followed by a materially incorrect or fabricated explanation, calculation, or code. No Score 2 responses occurred on factual questions. On false-premise questions, one Score 2 response occurred for Llama-3.1-8B-Instruct (1/30, 3.3%; Wilson 95% CI 0.6–16.7%) and none for Llama-3.3-70B-Versatile (0/30; Wilson 95% CI 0.0–11.4%). The findings are exploratory because the benchmark is small, only two models were tested, and the evaluation was conducted by a single researcher.