Back to AI News
arxiv

Definitions, Label Names, or Demonstrations? Auditing What Governs LLM Annotation

Researchers introduce a counterfactual audit that reassigns category definitions among output labels to test whether LLM annotations follow supplied definitions or preserve a dataset's original label mapping. Across three political-text datasets, models rarely follow reassigned definitions, and standard fine-tuning strengthens the original mapping without reliably improving definition adherence. In retrieval-augmented annotation, retrieving nearest training examples generally beats within-category retrieval despite using fewer demonstrations; on two tasks, direct k-nearest-neighbor classification recovers 74–93% of the improvement, and aligning demonstrations with reassigned definitions increases adherence to the new mapping.

AI Brief

It provides a concrete audit method and evidence that high annotation accuracy may reflect label or demonstration cues rather than adherence to researcher definitions.

arXiv:2611.09457v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly used to annotate social-science texts, but high accuracy does not establish that they apply researcher-supplied category definitions: predictions may instead reflect prior associations with category labels or labeled demonstrations. We introduce a counterfactual audit that reassigns definitions among output labels, distinguishing adherence to supplied definitions from preservation of a dataset's original mapping. Across three political-text datasets, models rarely follow reassigned definitions, and standard fine-tuning strengthens the original mapping without reliably increasing definition adherence. We then study retrieval-augmented annotation. Retrieving the nearest training examples generally outperforms retrieving examples separately within each category despite using fewer demonstrations. On the two tasks with substantial gains, direct \(k\)-nearest-neighbor classification recovers 74--93\% of the improvement, indicating that neighborhood labels provide much of the predictive signal. Relabeling fixed demonstrations redirects predictions, whereas aligning demonstrations with reassigned definitions substantially increases adherence to the new mapping. Thus, accuracy-improving interventions need not increase codebook adherence; definitions, category labels, demonstrations, retrieval, and training data should be aligned and validated jointly.
Be the first to rate this article.

Comments & Questions (0)

Captcha
Please be respectful — let's keep the conversation friendly.

No comments yet

Be the first to comment!