Abstract
Large Language Models (LLMs) are frequently criticized for "AI-splaining"—the generation of confident but factually incorrect statements. This position paper argues that for LLMs to become truly trustworthy, particularly in high-stakes domains like medicine or for low-resource languages, they must be engineered to cite the specific portions of their training data that support their outputs. By moving beyond post-hoc web searches toward internal attribution, the authors propose a roadmap for more transparent and verifiable artificial intelligence.
Introduction: The Stochastic Parrot Problem
The "Stochastic Parrot" phenomenon describes the tendency of LLMs to mimic patterns found in their training data without an underlying grasp of objective truth. This lack of grounding leads to hallucinations, where models generate plausible-sounding but false information. This issue is particularly acute in low-resource languages, where training data is often scarce, uncurated, or of lower quality.
While current tools like Perplexity AI or Retrieval-Augmented Generation (RAG) provide citations by searching the web after a query is processed, the authors argue this is an insufficient "external patch." These methods do not explain the model’s internal parameters or the specific data that influenced its response. To achieve genuine transparency, a model should be capable of pointing back to its own training set to justify its claims.
A Framework for Internal Attribution
As a position paper, this work proposes a conceptual framework and a roadmap for the NLP community rather than a new algorithm. The authors introduce several key concepts to guide the development of citation-aware models:
1. The "When to Cite" Taxonomy
Not every model output requires a citation. The authors propose a classification framework based on two primary questions:
- Is the source obvious? (e.g., in machine translation, the source is the input sentence itself).
- Is the text objective or subjective?
Based on this, they suggest a matrix for NLP tasks. For example, "Creative Writing" may only require citations occasionally to avoid plagiarism, whereas "Free-text Question Answering" should always require citations to ensure factuality.
2. STANCE and MultiSTANCE
The authors propose STANCE (Simultaneous Text ANd Citation gEneration), a task where models are trained to generate text and citations concurrently. This differs from post-hoc retrieval by making the citation an integral part of the generation process. This is extended into MultiSTANCE, which addresses "multihop" citations—cases where a single claim is supported by information distributed across multiple training documents.
3. Evaluating Quality with CUE
To measure success, the authors propose CUE (Citation qUality Evaluation). This metric would assess the quality of a source based on its publication date, credibility, and inherent bias, ensuring that models don't just cite any source, but the best source.
Citations as Interpretability
A core insight of the paper is that citations serve as a form of model interpretability. Knowing which training data triggered a specific response allows researchers and users to understand why a model reached a certain conclusion. This is especially critical for low-resource languages, where the web may not contain enough relevant factual sources for a standard search engine to find. In these cases, the model must rely on its internal training data, making the ability to verify that data even more vital.
Challenges and Ethical Considerations
The transition to citation-aware models is not without significant risks, which the authors candidly acknowledge:
- The "False Authority" Trap: There is a danger that users will be more likely to trust a hallucination if it is accompanied by a citation, even if that citation is irrelevant or incorrect.
- Readability vs. Verifiability: Excessive citation can clutter text, making it difficult for humans to read and process.
- Privacy and Data Leakage: If a model cites its training data directly, it may inadvertently leak sensitive information, such as private medical records or PII (Personally Identifiable Information) contained in the training set.
- Copyright and Licensing: While citations provide credit, they do not resolve the complex legal issues surrounding the use of copyrighted material for model training.
Conclusion
The authors call for a shift in the NLP community toward models that are "born" with the ability to justify their knowledge. By developing better subjectivity classification, ensuring citation-text correctness, and establishing protocols for source trustworthiness, the field can move toward a future where stochastic parrots finally tell us who they are imitating.
Summary updated