Bhavyajeet Singh

Large language models have shown remarkable progress in code generation and decision support, yet their ability to reliably follow instructions and integrate feedback remains limited-particularly in safety-critical domains. We plan to develop a framework for feedback-aligned instruction following, where models are trained and evaluated on their capacity to adapt to explicit guidance, iterative feedback, and corrective signals. Our approach seeks to improve both security and trustworthiness in Al outputs, with code generation serving as a primary test bed due to its high stakes in software safety and vulnerability prevention. Beyond coding, the framework can further generalizes to sensitive applications such as mental health support, where adherence to instructions and user intent is equally critical. By closing the loop between instructions, feedback, and generation, this work aims to advance the reliability and safety of AI systems across domains.