Abstract
Large language models (LLMs) exhibit remarkable text-generation capabilities, yet struggle with factual consistency, motivating growing interest in factuality verification. Existing factuality verification methods typically follow a Decompose-Then-Verify paradigm, which improves granularity but suffers from poor scalability and efficiency. We propose a novel Decompose-Embed-Interact paradigm that shifts factuality verification from costly text-level reasoning to efficient alignment in embedding space, effectively mitigating the scalability bottlenecks and computational inefficiencies inherent to prior approaches. While the proposed paradigm promises scalable verification, its implementation faces three practical challenges: efficient decomposition, factually faithful embedding, and accurate verification in embedding space. To address these challenges, we introduce E-Verify, a lightweight framework that resolves them through three specially designed modules, each aligned with a specific stage of the paradigm and designed to preserve scalability and efficiency. Experiments demonstrate that E-Verify significantly improves both decomposition and verification efficiency while maintaining competitive accuracy. These results confirm that the proposed paradigm enables scalable and fine-grained factuality verification with minimal performance trade-offs.
Introduction
Large Language Models (LLMs) are notorious for "hallucinating" facts—generating content that sounds plausible but is factually incorrect. To combat this, researchers have developed factuality verification systems. However, as LLM outputs get longer and knowledge bases grow, current verification methods are hitting a wall. This paper introduces E-Verify, a framework designed to make fact-checking both highly accurate and incredibly fast by shifting the computational burden from text-level reasoning to embedding-space alignment.
The Scalability Bottleneck
The current gold standard for fact-checking follows a Decompose-Then-Verify paradigm: break a generated text into "atomic facts" and then use a powerful model (like an LLM or an NLI model) to check each fact against reference documents. The primary bottleneck is the "Verify" step. Comparing every generated fact against every chunk of reference text leads to quadratic computational overhead ($O(N^2)$). For long documents or large knowledge bases, this becomes prohibitively expensive.
E-Verify breaks this scalability barrier by introducing the Decompose-Embed-Interact paradigm. This approach moves the heavy lifting from costly text-level reasoning to efficient vector-level operations in embedding space.
The E-Verify Framework
The E-Verify framework replaces text-to-text comparison with a three-stage vector-to-vector interaction process:
- Decomposition: Instead of relying on massive LLMs for fact extraction, E-Verify utilizes a fine-tuned Small Language Model (SLM), specifically a Qwen2.5-0.5B model. It processes text sentence-by-sentence to mitigate long-context hallucinations and maximize efficiency.
- Embedding: Atomic facts are converted into dense vectors. To capture subtle factual nuances, the authors introduce Pooling-based Multi-Head Attention (PMA). This creates "multi-view" embeddings—specifically "Direct" and "Diff" embeddings—that preserve richer contextual information than standard pooling methods.
- Interaction: Rather than using a slow transformer-based cross-encoder, E-Verify employs a lightweight Multi-Feature Interaction Module (MFIM). It calculates two key features in embedding space: a Pairwise Feature for surface-level matches and a Discrepancy Feature that uses vector subtraction to model the directional nature of factual entailment.
Key Results and Insights
Experimental evaluation across benchmarks like wiki-bio-hallu and CNN demonstrates that E-Verify achieves massive efficiency gains without sacrificing performance:
- Efficiency: E-Verify is 60x faster at fact decomposition than GPT-4o and 22x to 46x faster at verification than state-of-the-art NLI-based models like MiniCheck.
- Scalability: While NLI-based methods show quadratic growth in processing time, E-Verify's growth is nearly linear, making it suitable for large-scale applications.
- Accuracy: Despite its smaller footprint, E-Verify's accuracy remains competitive with traditional NLI baselines and significantly outperforms standard embedding models like BGE.
Conclusion
E-Verify represents a paradigm shift in factuality verification. By moving from text-level reasoning to embedding-space interaction, it addresses the critical scalability challenges of modern LLM evaluation. This approach enables real-time, fine-grained fact-checking for long-form content, providing a scalable solution for ensuring the reliability of generated text.
Summary updated