Real-Time Detection of Hallucinated Entities in Long-Form Generation (www.hallucination-probes.com)

Researchers have introduced a novel, scalable method for real-time detection of hallucinated entities—such as fabricated names, dates, and citations—in long-form text generated by large language models (LLMs). Unlike existing approaches that rely on costly and slow external verification, this technique operates token-by-token during generation, enabling streaming hallucination detection especially suited for complex, multi-paragraph outputs. By focusing on entity-level hallucinations rather than entire claims, the system exploits clear token boundaries, training lightweight linear probes on hidden model activations to accurately flag hallucinated content as it emerges. The team created a large annotated dataset, LongFact++, leveraging web-search-enabled frontier LLMs to verify entities within model completions, providing precise token-level labels for training. Across multiple LLM families—including 70B parameter models like Llama 3.3—the probes consistently outperformed traditional uncertainty-based baselines and expensive methods such as semantic entropy, achieving AUC scores above 0.85 in long-form tasks and over 0.96 in short-form QA settings. Remarkably, the probes also generalized to out-of-distribution tasks like mathematical reasoning. Additional experiments demonstrated cross-model generalization and showed that probes trained on long-form data transfer well to short-form detection, but not vice versa. Beyond improved accuracy, this real-time detection enables new use cases like selective answering—abstaining from generating potentially hallucinated content mid-generation to increase reliability. This work marks a significant step toward practical, low-latency hallucination monitoring in LLMs, addressing a critical challenge for deploying AI safely in high-stakes domains such as healthcare and law.