Fine-grained citations may harm AI model performance, research finds

Researchers examining citation granularity in AI-generated text have discovered a counterintuitive finding: the finest level of citation detail isn't always the best approach. Their analysis of models ranging from 8 billion to 120 billion parameters reveals that restricting citations to individual sentences often comes at a substantial performance cost.

The Performance Trade-Off#

The study measured the impact of different citation granularities — whether sources are cited at the sentence, paragraph, or document level — across multiple model scales. The results show that enforcing sentence-level citations alone causes performance losses of 2 to 97 percent relative to the optimal granularity for each task, with some individual cases showing degradation as high as 338 percent. Notably, this performance decline occurred even as overall answer correctness remained nearly stable, dropping by no more than 2.3 percent or rising by up to 4.4 percent.

Why Intermediate Granularity Works Best#

The research reveals a consistent pattern across all tested models: attribution quality peaks at intermediate (paragraph-level) granularities. The explanation lies in semantic structure. Fine-grained citations that break claims into single sentences appear to sever the connections between multiple related pieces of evidence that together form a complete idea. Conversely, excessively broad citations at the document level introduce too much irrelevant context, creating confusion rather than clarity.

This trade-off becomes more pronounced with larger models. When evidence for a claim spans multiple sentences, forcing sentence-level citations disproportionately penalizes bigger models, which are optimized to synthesize information across sentence boundaries.

Challenging Human-Centered Assumptions#

The findings question a foundational premise of current citation standards: that a single sentence constitutes sufficient evidence on its own. The researchers argue that this assumption often fails in practice. They contend that "standards fixed for human verifiability may paradoxically degrade the very attribution they aim to ensure." Rather than adhering to fixed conventions, effective attribution requires matching citation scope to how the model actually processes and grounds information.

The implication is striking: optimal granularity is a property of the model's capabilities, not an immutable principle of citation design.