Research map / Large language models / Language, safety and interpretability
ABCD: All Biases Come Disguised
Mateusz Nowak, Xavier Cadet, Peter Chin
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 |
|---|---|---|---|
| replace letter labels and letter prediction (S&L) with dash labels and matched full-text answers (M&D) | 13 LLMs · five MCQ benchmarks | accuracy variance across permutations: variance ratio geometric mean 0.33 | Improved |
| use M&D instead of S&L | 13 LLMs · NonsenseQA, 5-shot | median accuracy: per-model median reductions 12.5% to 52% | Improved |
| use M&D instead of S&L | 13 LLMs · CSQA | variance: variance ratio 0.11 | Improved |
| use M&D instead of S&L | Llama-3.1-8b-instruct · CSQA | variance: variance ratio 1.688 | Hurt |
| use M&D instead of S&L | 13 LLMs · ARC | variance: 11 of 13 models lower variance | Improved |
| use M&D instead of S&L | 13 LLMs · GPQA | variance: 9 of 13 models lower variance | Improved |
| use M&D instead of S&L | 13 LLMs · MMLU-Pro | variance: 7 of 13 cases lower variance; ratio 0.99 | No clear effect |
| use M&D instead of S&L | 13 LLMs · INCLUDE (es, fr, it, de) | variance: variance ratio 0.23 | Improved |
| use M&D instead of S&L | 13 LLMs · SCORE robustness on MMLU-Pro and ARC | SCORE | Hurt |
| use M&D instead of S&L when few-shot distribution is unchanged | 13 LLMs · GPQA | SCORE: 10 of 13 models improved | Improved |
| vary sentence similarity model and similarity function (ablation) | M&D matching · benchmarks | accuracy | No clear effect |
| use homogeneous option labels instead of heterogeneous labels (ablation) | M&D · benchmarks | evaluation stability | Improved |
Closest papers by meaning
- B-score: Detecting biases in large language models using response history (ICML 2025 · similarity 0.89)
- Large Language Models Are Not Robust Multiple Choice Selectors (ICLR 2024 · similarity 0.89)
- Rethinking Generative Large Language Model Evaluation for Semantic Comprehension (ICML 2024 · similarity 0.89)
- Answer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questions (ICLR 2025 · similarity 0.88)
- Evaluating language models as risk scores (NeurIPS 2024 · similarity 0.88)
- Representation Consistency for Accurate and Coherent LLM Answer Aggregation (NeurIPS 2025 · similarity 0.88)
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.
