Enterprises increasingly would like to use Large Language Models (LLMs) to enhance operational efficiency, yet verifying their accuracy remains a critical challenge. Our LLM Explainer helps build trust and transparency by linking AI-generated answers to their exact source sentences. This allows fast and efficient verification of AI output, particularly suitable for high-risk domains. LLM Explainer achieves this through automated data generation and the training of efficient attribution models. Evaluation shows it obtains superior accuracy while operating at 1/70th of the computational cost of larger models with lower accuracy. For decision-makers, it transforms the use of LLMs from a potential vulnerability into a strategic capability that strengthens confidence in mission-critical intelligent systems.

1. Introduction

Large Language Models (LLMs) are rapidly becoming core components of enterprise operations, driving automation, insight generation, and customer engagement. However, their growing adoption also amplifies critical concerns around reliability, transparency, and compliance. One of the most pressing challenges is the prevention of inaccurate or misleading outputs—often referred to as hallucinations—which can erode trust, disrupt workflows, and expose organizations to regulatory or reputational risk. As enterprises seek to leverage LLMs for high-stakes decision-making, ensuring the verifiability of AI-generated content has become essential to maintaining accountability, operational integrity, and stakeholder confidence.

Our technology addresses this challenge by linking LLM answers to specific source sentences. This enable the efficient verification of LLM output and is an essential capability for organizations requiring transparency, auditability, and reliability in AI-driven outputs.
 
Our novel approach, the LLM Explainer, leverages advanced LLMs to automatically generate high-quality synthetic training data that enables smaller, efficient models to perform real-time, sentence-level context attribution with high accuracy¹⁾. Context attribution is the task of linking an LLM-generated sentence to a source sentence that supports or refutes the LLM-generated sentence. Unlike existing methods that rely on coarse document-level evidence or computationally expensive inference, LLM Explainer delivers fast and precise attribution, significantly reducing the need for costly human annotation and enabling scalable deployment across diverse domains. For an example, see the figure below.
 
For NEC’s customers—spanning government agencies, telecommunications providers, financial institutions, and transportation sectors—this technology offers concrete business benefits: enhanced trustworthiness of AI-generated content, improved operational efficiency through faster and more accurate verification of answers, and enabling compliance with stringent regulatory requirements for transparency and accountability. Our user studies demonstrate that LLM Explainer's attribution models help end users verify AI outputs more quickly and accurately.
 
By using LLM Explainer for context attribution, users gain a robust, scalable tool that strengthens the reliability of automated question answering and supports mission-critical decision-making processes. This advancement enables to deliver secure, transparent, and trustworthy AI technologies that meet the complex needs of regulated industries and public sector organizations worldwide.

Fig.1 description: Given a question, an LLM-generated answer, and context (from human input or retrieval), the model identifies supporting sentences within the context. Our user study shows that presenting these supporting sentences helps users verify LLM answers more quickly and accurately.

2. System Overview

Our solution presents an innovative automated source attribution system that enables real-time verification of AI-generated text while maintaining operational efficiency.  The system provides immediate traceability by automatically identifying and citing the specific source materials that support each AI-generated response.

The technology operates through a post-verification approach that analyzes AI-generated content after generation, eliminating the need to modify existing AI systems or workflows. The core architecture consists of three integrated components:

3. LLM Explainer

Existing source attribution systems typically require extensive training data to accurately identify which parts of reference documents support AI-generated responses. Traditional approaches demand costly human annotation processes, creating implementation barriers for organizations seeking to deploy verification systems across large-scale operations. Our solution addresses this challenge through automated training data generation that eliminates human annotation requirements while producing superior results.

Source attribution identifies which portions of a trusted set of documents support specific AI-generated question-answer pairs. In formal terms, given a query, its AI-generated response, and organizational context documents consisting of multiple sections, the system must identify the specific document sections that fully support (or refute) the generated response. To train efficient verification models without expensive human oversight, we developed an automated data generation approach using advanced AI systems.

We implement two distinct methods for automated training data creation. The baseline approach uses a discriminative method: existing question-answer pairs and their organizational context are processed by an AI system to identify supporting document sections, which then train smaller verification models through knowledge transfer. In contrast, our proposed method employs a generative approach: selected document sections are used by an AI system to generate question-answer pairs that are fully supported by these sections. This approach better utilizes AI systems' natural text generation capabilities while ensuring clear evidence trails in the training data.

3.1 Training Setup

Data Generation for training LLM Explainer operates through three integrated components: context selection, question-answer generation, and distractor integration. This comprehensive approach ensures training data reflects real-world enterprise scenarios where verification systems must distinguish between relevant and irrelevant information sources.

3.2 Strategic Advantages

Our approach delivers three critical advantages for enterprise AI verification implementations:

Top of Fig.2: The baseline method for synthetic attribution data generation. Given context and question-answer pairs, we prompt an LLM to identify supporting sentences, which are then used to train a smaller attribution model. However, this discriminative approach may yield noisy training data as LLMs are less suited for classification tasks. Bottom of Fig. 2: Our data generation pipeline leverages LLMs’ generative strengths through four steps:

4. Application and Evaluation of LLM Explainer

We conducted comprehensive validation of LLM Explainer’s source attribution technology across multiple scenarios to demonstrate its practical utility and performance advantages. Our evaluation compares our LLM Explainer models against two primary alternatives:

4.1 Evaluation Framework and Metrics

Enterprise AI verification systems must balance three critical performance factors: accuracy in identifying correct sources, completeness in finding all relevant evidence, and overall reliability. We measure system performance using standard metrics adapted for source attribution:

4.2 Results of Automatic Evaluation

We describe the results per task.

4.3 Results of User Study

4.4 Enterprise Value of LLM Explainer

The validation results translate directly to measurable business benefits for organizations implementing source attribution technology. LLM Explainer delivers high accuracy while reducing computational costs by over 98% compared to large-scale alternatives, enabling deployment on standard business hardware without specialized infrastructure. The system's ability to automatically generate training data eliminates the months-long human annotation processes that typically cost a significant amount of money, while maintaining superior performance across diverse enterprise scenarios from regulatory compliance to technical documentation verification.

5. Conclusion

Our automated source attribution technology via LLM Explainer addresses a critical challenge facing enterprise AI deployments: enabling organizations to harness AI productivity benefits while maintaining the accountability and verification standards required for mission-critical operations. Through the development of our automated training data generation and efficient LLM Explainer models, we have demonstrated a practical solution that eliminates the cost and complexity barriers traditionally associated with AI verification systems. The technology delivers superior performance compared to both expensive large-scale AI alternatives and traditional human-annotated approaches, while operating efficiently on standard enterprise hardware.

LLM Explainer consistently outperformed zero-shot systems that are 70 times larger, while demonstrating robust cross-domain generalization essential for organizations with varied content types and business processes. The system's ability to automatically generate high-quality training data eliminates dependence on costly human annotation processes, enabling rapid deployment across new departments or content domains without extensive preparation periods. These capabilities lower the barrier of adopting AI systems in regulated environments, customer-facing applications, and mission-critical workflows where verification and accountability are paramount.

References

Authors’ Profiles