Scalable Token-Level Hallucination Detection in Large Language Models

2026-05-12

AI Summary

This work addresses the challenge of fine-grained hallucinations in large language models during complex reasoning, a problem inadequately tackled by existing detection methods due to their reliance on coarse-grained step segmentation and limited scalability. The authors propose TokenHD, a novel framework that enables token-level hallucination detection without requiring predefined step boundaries. TokenHD leverages a scalable hallucination data synthesis engine to generate large-scale annotated training data and employs an importance-weighted training strategy to end-to-end identify hallucinated tokens in free-form text. Remarkably, a lightweight detector with only 0.6B parameters outperforms reasoning models as large as 32B and demonstrates consistently improved performance across model scales from 0.6B to 8B. The framework exhibits strong generalization and scalability across diverse domains.

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities, but they still frequently produce hallucinations. These hallucinations are difficult to detect in reasoning-intensive tasks, where the content appears coherent but contains errors like logical flaws and unreliable intermediate results. While step-level analysis is commonly used to detect internal hallucinations, it suffers from limited granularity and poor scalability due to its reliance on step segmentation. To address these limitations, we propose TokenHD, a holistic pipeline for training token-level hallucination detectors. Specifically, TokenHD consists of a scalable data engine for synthesizing large-scale hallucination annotations along with a training recipe featuring an importance-weighted strategy for robust model training. To systematically assess the detection performance, we also provide a rigorous evaluation protocol. Through training within TokenHD, our detector operates directly on free-form text to identify hallucinations, eliminating the need for predefined step segmentation or additional text reformatting. Our experiments show that even a small detector (0.6B) achieves substantial performance gains after training, surpassing much larger reasoning models (e.g., QwQ-32B), and detection performance scales consistently with model size from 0.6B to 8B. Finally, we show that our detector can generalize well across diverse practical scenarios and explore strategies to further enhance its cross-domain generalization capability.

Problem

• hallucination detection
• large language models
• token-level
• reasoning tasks
• scalability

Innovation

• token-level hallucination detection
• scalable data engine
• importance-weighted training
• step-free detection
• cross-domain generalization