Research map / Large language models / Language, safety and interpretability
On the Limits of LLM Adaptability: Impact of Model-Internalized Priors on Annotation Task Performance
Etienne Casanova, Rafal Kocielnik, R. Michael Alvarez
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the aligned task definition to the prompt | 9 LLMs · Toxicity annotation (5 datasets) | accuracy (%): 80.3 -> 82.0 | Improved |
| add few-shot examples to the prompt | 9 LLMs · Toxicity annotation (5 datasets) | accuracy (%): 80.3 -> 81.5 | Improved |
| add the definition to few-shot prompting | 9 LLMs · Toxicity annotation (5 datasets) | accuracy (%): 81.5 -> 81.6 | No clear effect |
| apply DSPy automated prompt optimization | 9 LLMs · Toxicity annotation (5 datasets) | accuracy (%): 80.3 -> 80.2 | No clear effect |
| apply DSPy optimization | Llama-3.1-8B · Toxicity annotation | accuracy (%): 76.7 -> 72.3 | Hurt |
| apply DSPy optimization | Llama-3.3-70B · Toxicity annotation | accuracy (%): 80.5 -> 77.5 | Hurt |
| use misaligned definitions instead of the aligned definition | 9 LLMs · Toxicity annotation (5 datasets) | accuracy (%): 82.0 -> 80.0 | Hurt |
| swap aligned definition for Gaming toxicity definition | 9 LLMs · Toxicity annotation | accuracy (%): 82.0 -> 76.4 | Hurt |
| swap aligned definition for Fox News hate speech definition | 9 LLMs · Toxicity annotation | accuracy (%): 82.0 -> 82.6 | Improved |
| increase zero-shot confidence by one SD | 9 LLMs · Toxicity annotation, error correction | odds ratio of rescue: OR=0.84 | Hurt |
| add definition-based prompting (high-confidence zero-shot errors vs moderate) | 9 LLMs · Toxicity annotation, rescue by confidence | rescue probability (%): 51.8% -> 20.8% | Hurt |
| use larger model (Llama-3.1-70B instead of Mistral-Small-24B) | LLM annotators · Toxicity annotation | rescue rate (%): 44.2% -> 31.3% | Hurt |
Closest papers by meaning
- Task Contamination: Language Models May Not Be Few-Shot Anymore (AAAI 2024 · similarity 0.90)
- Gaining Wisdom from Setbacks: Aligning Large Language Models via Mistake Analysis (ICLR 2024 · similarity 0.89)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.89)
- Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting (ICLR 2024 · similarity 0.89)
- Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.89)
- Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning (ICLR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
