{
  "id": 3406,
  "url": "https://arxiv.org/abs/2606.00467v1",
  "title": "On the Limits of LLM Adaptability: Impact of Model-Internalized Priors on Annotation Task Performance",
  "summary": "Large Language Models (LLMs) are increasingly used for zero-shot annotation and LLM-as-a-judge tasks, yet their reliability hinges on how model-internalized priors interact with user-provided instructions. We investigate three dimensions of this interaction: (1) how an LLM's familiarity with data and task definitions affects performance, (2) the extent to which additional information in prompts can correct zero-shot errors (\"decision stickiness\"), and (3) model susceptibility to misaligned task ",
  "authors": "Etienne Casanova, Rafal Kocielnik, R. Michael Alvarez",
  "category": "research",
  "topics": "safety-alignment,finance-investment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-30T01:21:14.000Z",
  "fetched_at": "2026-07-14T16:30:14.369Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/3406",
  "original_url": "https://arxiv.org/abs/2606.00467v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}