A novel technique called Selective Prompt Anchoring (SPA) that improves LLMs' code generation accuracy by preventing attention drift during generation.

• LLMs tend to pay less attention to user prompts as they generate more code tokens
• This attention dilution leads to code generation errors and misalignment with user intent
• SPA technique helps maintain focus on critical prompt elements throughout the generation process
• Particularly valuable for complex programming tasks where sustained attention to requirements is essential

This research addresses a fundamental engineering challenge in AI-assisted programming, enabling more reliable code generation tools that better align with developer requirements.

Selective Prompt Anchoring for Code Generation