Annotation Entropy Predicts Per-Example Learning Dynamics in LoRA Fine-Tuning

A new research paper by Brady Steele, titled "Annotation Entropy Predicts Per-Example Learning Dynamics in LoRA Fine-Tuning," reveals a critical and counterintuitive flaw in a popular AI training method. The study demonstrates that when using LoRA (Low-Rank Adaptation) for parameter-efficient fine-tuning, models don't just learn slowly from ambiguous data—they actively get worse. On examples with high annotator disagreement (measured as 'annotation entropy' from datasets like ChaosNLI with 100 labels per item), the model's loss actually increases during training, a phenomenon the authors term 'un-learning.' This pattern was absent in traditional full fine-tuning and was observed consistently across all six tested models, including four encoder and two decoder-only architectures.

The research, based on analyzing the SNLI and MNLI datasets, found a positive correlation between annotation entropy and the area under the loss curve (AULC) in all 25 experimental conditions. The strength of this correlation varied (Spearman ρ = 0.06 to 0.43), with decoder-only models showing stronger effects than encoders at matched LoRA rank. The findings survived rigorous statistical controls for partial correlation and replicated across different random seeds and datasets. A preliminary noise-injection experiment further supported the conclusion that LoRA is uniquely sensitive to label noise, causing it to deteriorate performance on contested examples rather than improving or ignoring them.

This discovery has significant implications for how practitioners approach fine-tuning, especially for applications dealing with subjective or ambiguous data. It suggests that the choice between full fine-tuning and parameter-efficient methods like LoRA isn't just about computational cost—it fundamentally changes what the model learns. The paper provides a new diagnostic tool (annotation entropy) for predicting which examples will be problematic and opens the door for developing more robust fine-tuning algorithms that can handle real-world data ambiguity.