Learning to Say "I Don't Know": A Vision for Abstention in Large Language Models
Tapomoy Adhikari · 2025
Large Language Models (LLMs) have achieved remarkable progress in open-domain language generation, but they remain prone to producing fluent yet incorrect information-commonly known as hallucinations. In domains where factual accuracy is critical, such behavior undermines reliability, propagates misinformation, and erodes trust. We propose integrating an explicit abstention capability into LLMs: the deliberate refusal to answer when the model lacks sufficient confidence or supporting evidence. Unlike classification tasks, where decision boundaries enable probabilistic abstention, LLMs operate in an unconstrained generative space where token likelihoods do not reliably reflect semantic correctness. This vision paper outlines the rationale, benefits, and probable approaches for realizing abstention in LLMs, positioning it as a cornerstone for safe and trustworthy AI.