Diagnosing Failures in Large Language Models' Answers: Integrating Error Attribution into Evaluation Framework
Introduces MisAttribution Framework and AttriData dataset; trains MisAttributionLLM for multi-task error evaluation with high correlation (0.935).
Key Findings
Methodology
The paper develops a comprehensive hierarchical misattribution framework with 6 primary and 15 secondary categories, covering response quality, safety, reasoning, and knowledge. Using expert-annotated AttriData, it trains a 7B parameter Qwen-based model via supervised fine-tuning, employing multi-task learning to simultaneously generate scores, error types, and feedback. The annotation process involves multiple rounds of validation, achieving Fleiss’ kappa of 0.875, ensuring high consistency. The model’s performance is validated through correlation metrics (Pearson 0.935), error detection, multi-classification, and human preference tests, demonstrating robustness and superior accuracy compared to GPT-4 and GPT-3.5.
Key Results
- MisAttributionLLM achieves a Pearson correlation of 0.935 with human scores, surpassing GPT-4 (0.817) and GPT-3.5 (0.411). Its micro-F1 score for error attribution reaches 0.829, outperforming several open- and closed-source models.
- In error detection, the model attains the highest accuracy, with strong recall, indicating effective identification of errors. Multi-label classification results show high precision and robustness in complex error scenarios.
- Human preference tests reveal that feedback generated by MisAttributionLLM is preferred over GPT-4 in 60.41% of cases and over GPT-3.5 in 85.20%, confirming high-quality feedback generation.
Significance
This work addresses a critical gap in LLM evaluation by enabling detailed error attribution, which is essential for model debugging, safety, and user trust. The hierarchical framework and high-quality dataset facilitate precise diagnosis of model weaknesses, guiding targeted improvements. The open-source MisAttributionLLM provides a practical tool for researchers and developers to implement fine-grained evaluation, fostering transparency and robustness in LLM deployment. Overall, it advances the state of the art in automated model assessment, with broad implications for AI safety, interpretability, and industrial application.
Technical Contribution
The paper introduces a multi-level misattribution framework, combining empirical analysis and expert annotation, to classify errors comprehensively. It constructs AttriData, a large-scale, high-quality dataset for supervised training. The core technical innovation is the adaptation of multi-task learning in a 7B parameter model, enabling simultaneous scoring, error attribution, and feedback generation. The evaluation metrics demonstrate the model’s high correlation with human judgments and superior error classification accuracy, setting new benchmarks for automated error analysis in LLMs.
Novelty
This is the first work to establish a detailed hierarchical error attribution framework integrated with a large-scale annotated dataset and to train an open-source LLM capable of multi-task evaluation. Unlike prior models that focus solely on scoring or feedback, MisAttributionLLM provides comprehensive error diagnosis, including specific error types and explanations, representing a significant step forward in interpretability and safety in LLM evaluation.
Limitations
- The model’s performance may decline in unseen or highly complex error scenarios, especially in domains with limited training data. Its effectiveness across languages beyond Chinese and English remains untested.
- High computational costs associated with training and inference may limit deployment in resource-constrained environments. Further optimization is needed for real-time applications.
- Despite high annotation consistency, some subtle errors may still be misclassified, indicating room for improving the granularity and scope of the error categories.
Future Work
Future research will expand the hierarchical error categories, incorporate multi-modal data (e.g., images, videos), and enhance cross-lingual transferability. Integrating user feedback loops for continuous model refinement and developing lightweight versions for deployment in real-time systems are also promising directions. Additionally, exploring reinforcement learning strategies to optimize feedback quality could further improve model robustness and safety.
AI Executive Summary
The rapid proliferation of large language models (LLMs) has revolutionized natural language processing, yet their evaluation remains a complex challenge. Traditional metrics like BLEU or ROUGE, while useful, do not capture the nuanced errors that can compromise safety, reliability, and user trust. Recognizing this gap, the present work introduces a hierarchical misattribution framework that categorizes errors into six primary and fifteen secondary types, covering critical aspects such as response quality, safety, reasoning, and knowledge accuracy.
Building on this framework, the authors curated AttriData, a large-scale, expert-annotated dataset encompassing over 21,000 samples, with detailed labels for error types, scores, and feedback. This dataset serves as the foundation for supervised fine-tuning of a 7B parameter Qwen-based model, resulting in MisAttributionLLM—a multi-task model capable of scoring responses, diagnosing specific errors, and generating explanatory feedback.
Extensive experiments demonstrate the model’s high correlation with human judgments (Pearson 0.935), outperforming GPT-4 and GPT-3.5 in error attribution accuracy and feedback quality. The model’s error detection and classification capabilities are validated through metrics like micro-F1 (0.829) and human preference tests, where it is favored over leading closed-source models.
This research significantly advances automated model evaluation, providing a transparent, detailed, and scalable approach to diagnosing model weaknesses. Its implications extend to improving model safety, interpretability, and user trust, especially in high-stakes applications. Future directions include expanding hierarchical error categories, enhancing multi-modal capabilities, and optimizing for real-time deployment, promising a more robust and accountable AI ecosystem.
Deep Dive
Abstract
With the widespread application of Large Language Models (LLMs) in various tasks, the mainstream LLM platforms generate massive user-model interactions daily. In order to efficiently analyze the performance of models and diagnose failures in their answers, it is essential to develop an automated framework to systematically categorize and attribute errors. However, existing evaluation models lack error attribution capability. In this work, we establish a comprehensive Misattribution Framework with 6 primary and 15 secondary categories to facilitate in-depth analysis. Based on this framework, we present AttriData, a dataset specifically designed for error attribution, encompassing misattribution, along with the corresponding scores and feedback. We also propose MisAttributionLLM, a fine-tuned model on AttriData, which is the first general-purpose judge model capable of simultaneously generating score, misattribution, and feedback. Extensive experiments and analyses are conducted to confirm the effectiveness and robustness of our proposed method.