TokenHD: A New Fine-Grained Detection Method for Hallucinations in Large Language Models

TokenHD proposes a new token-level method for detecting hallucinations in large language models. By using a scalable data synthesis engine and an importance-weighted training strategy, it addresses the limitations of existing step-level detection methods, such as restricted granularity and poor scalability. Experiments show that even a small model with only 0.6B parameters can outperform the reasoning capabilities of large models with 32B parameters, achieving excellent performance in hallucination detection tasks and providing more precise solutions for scenarios like AI safety and content moderation.

Large language models tend to generate "hallucinations" (plausible but incorrect information) when producing content, which poses a trust barrier in professional scenarios like healthcare, law, and academia. Existing detection methods rely on step-level analysis and have two major flaws: first, restricted granularity—they can only locate steps rather than specific words; second, poor scalability—they require predefined rules or additional models, increasing complexity and overhead.

TokenHD adopts a token-level detection paradigm, with the core insight that hallucinations often stem from deviations in key tokens. Its architecture includes three key components: 1. A scalable data synthesis engine that automatically constructs training samples containing hallucination patterns, eliminating reliance on manual annotation; 2. An importance-weighted training strategy that makes the model pay more attention to high-risk tokens such as numerical values and proper nouns; 3. A systematic evaluation protocol covering metrics like hallucination patterns, cross-domain generalization, and detection latency.

TokenHD abandons predefined step segmentation and uses end-to-end detection, directly outputting the hallucination probability of each token. Its advantages include no need for text reorganization, ability to handle arbitrary free text, and precise localization of word-level errors. Experiments show that the 0.6B-parameter detector outperforms large models with 32B parameters (e.g., QwQ-32B), and its performance improves steadily as the model scale increases from 0.6B to 8B, making it flexible to adapt to resource-constrained scenarios.

On standard hallucination detection datasets, TokenHD shows significant improvements over baseline methods, especially in complex scenarios like numerical reasoning, fact-checking, and logical consistency. Cross-domain generalization tests demonstrate good adaptability, thanks to diverse synthetic samples and the weighted strategy. Detection latency is controlled at the millisecond level, meeting real-time interaction requirements and outperforming baseline methods with second-level latency.

TokenHD's fine-grained detection capability can be applied in: 1. Content moderation: Marking problematic parts in AI-generated content to avoid over-censorship; 2. Educational assistance: Identifying potential errors in teaching materials for teachers to verify; 3. Enterprise knowledge management: Serving as a security layer for RAG systems to detect inconsistencies between generated content and source documents, preventing fabricated information.

TokenHD has limitations: 1. The detector may make misjudgments, requiring optimization of the balance between recall and false positive rates; 2. It is difficult to capture overall logical fallacies (issues not related to individual tokens); 3. It only targets the text modality. Future directions include: reducing false positive rates, integrating high-level semantic analysis, and extending to multi-modal detection.

TokenHD represents an important advancement in the field of hallucination detection, proving that the fine-grained end-to-end paradigm can surpass traditional step-level methods. Through sophisticated data synthesis and training design, small specialized models can outperform large models in specific tasks, providing insights for the efficient use of model scale while offering practical AI safety solutions.