Refusal-Aware Instruction Tuning through Gradient-Guided Optimization and Knowledge-Flow Calibration for Hallucination Mitigation under Uncertain Knowledge Boundaries and Conflicting Evidence in Large Language Models

Abstract

The rapid deployment of large language models across diverse domains has exposed significant vulnerabilities related to factuality and reliability, particularly manifesting as hallucinations when models encounter queries beyond their intrinsic knowledge boundaries or process conflicting evidence. Traditional instruction tuning predominantly encourages compliance and response generation, inadvertently conditioning models to fabricate answers rather than accurately assessing and communicating their epistemic limitations. This paper introduces Refusal-Aware Instruction Tuning, a novel framework designed to mitigate hallucinations by explicitly training large language models to recognize their knowledge boundaries and refuse unanswerable queries. The proposed architecture integrates Gradient-Guided Optimization, which leverages internal gradient signals to detect regions of high epistemic uncertainty, and Knowledge-Flow Calibration, a mechanism that modulates attention pathways to resolve conflicting contextual evidence. By systematically mapping the internal latent representations of uncertainty, the model learns to output calibrated refusals without degrading its performance on answerable tasks. Extensive empirical evaluations across multiple complex reasoning and factuality benchmarks demonstrate that the proposed methodology significantly reduces hallucination rates while maintaining high utility. The experimental results establish that integrating refusal-awareness into the instruction tuning pipeline constitutes a critical step toward developing safe, reliable, and trustworthy artificial intelligence systems capable of navigating uncertain informational environments.