Skip to content
Modelwire
Subscribe

LLMs ignore definitions when annotating social science concepts

Source published ·Modelwire updated

Original coverage: arXiv cs.CL ↗·How Modelwire adds context

The development

Researchers have exposed a critical vulnerability in how LLMs perform social science annotation: models show high sensitivity to prompt-based definitions, yet may ignore them entirely in favor of learned biases. By annotating identical texts with competing definitions of concepts like hate speech and fake news, the study reveals whether LLMs actually follow instructions or default to pretraining. This finding directly challenges the reliability of LLM-assisted labeling pipelines in academic research and raises questions about reproducibility and definition-agnostic model behavior across subjective classification tasks.

Modelwire’s AI-generated summary of coverage from arXiv cs.CL.

Modelwire analysis

Explainer

Our AI-generated reading of the wider context and the next developments to watch.

The study isolates a specific failure mode: LLMs don't just vary in output, they actively ignore definitional instructions when those definitions conflict with learned patterns. This is distinct from general inconsistency.

This directly extends the robustness framework from the November 4th replicability paper, which showed that measurement choices can flip statistical conclusions. That work proposed standardized prompt protocols as a fix. This new study suggests the fix may be incomplete: even with explicit definitions in prompts, models revert to pretraining biases. The November 8th system prompts paper reinforces this pattern, showing that prompt framing drives behavior independent of actual discriminability. Together, these findings suggest that social scientists relying on LLM classifiers face a deeper problem than prompt variation alone: the models may not be following instructions at all, regardless of how carefully those instructions are written.

If researchers can show that fine-tuning on definition-aligned examples reduces sensitivity to competing definitions more than prompt engineering alone, that would indicate the problem is solvable within current workflows. If sensitivity persists across fine-tuned models, it signals that LLM-based annotation pipelines need human validation layers before publication, not just robustness checks.

This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error

Coverage behind this analysis

These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.

  1. ·arXiv cs.CL

    LLM output variance can reverse research findings, new reproducibility framework shows

    A new framework tackles a foundational problem in LLM-driven research: probabilistic output variation that can flip statistical conclusions. Rather than demanding identical model outputs, the work proposes measuring robustness across result distributions and introduces standardized prompt protocols to diagnose systematic model errors. The research demonstrates that measurement choices in LLM-based studies can reverse regression coefficients,…

    Read Modelwire coverage →Original source ↗

MentionsLLM

MW

How this coverage is produced

Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Measuring Definition Sensitivity in LLM Annotation of Social Science Concepts”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.

More to explore

System prompts drive LLM safety more than actual safety reasoning

Frontier LLMs resist most cognitive biases but fail on risk anchoring

Concept probe accuracy masks hidden bias from correlated attributes

LLMs ignore definitions when annotating social science concepts · Modelwire