Enhancing Trust in Language Model-Based Code Optimization through RLHF: A Research Design

AI code optimization tools can produce unreliable outputs: The paper identifies "hallucinations" (incorrect or fabricated responses) as a significant challenge when using language models for code optimization, highlighting that current AI systems cannot be blindly trusted with critical software engineering tasks.

Human feedback is essential for trustworthy AI: The research proposes using Reinforcement Learning from Human Feedback (RLHF) to improve language models' reliability, suggesting that incorporating human oversight and correction is a practical pathway to making AI-driven code tools more dependable.

Trust in AI requires intentional design: Rather than assuming language models are inherently trustworthy, this work emphasizes that building reliable AI systems for software engineering demands deliberate research and development focused specifically on trustworthiness and human-centric approaches.

Software engineering is a critical domain for responsible AI: The focus on code optimization underscores that AI systems used in software development warrant special attention to reliability, since errors in code could have downstream consequences for end users and systems.