Research map / Large language models / Retrieval and knowledge
Does Question Really Matter? The Attribution of Answer Bias in LLM Evaluation
Boxi Cao, Ruotong Pan, Hongyu Lin, Xianpei Han, Le Sun
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 |
|---|---|---|---|
| shuffle option order in option-only evaluation (ablation) | 17 LLMs · ARC-challenge option-only | accuracy: 38.37% -> 38.19% | No clear effect |
| replace wrong options with random wrong options from other questions (ablation) | Qwen2-72B · MMLU option-only | accuracy: 48.8% -> 35.3% | Hurt |
| train on test data (more contaminated samples) (ablation) | base LLMs · MMLU/CMMLU/HellaSwag option-only | option-only accuracy: Pearson r 0.935, 0.991, 0.925 | Improved |
| train on contaminated subjects vs uncontaminated control subjects (ablation) | LLaMA-3-8B · CMMLU | option-only accuracy: contaminated 62.9 vs control 34.4 | Improved |
| use OPD instead of BLD for contamination detection | LLaMA-2-7B contaminated with HellaSwag · contamination detection | F1: 0.558 -> 0.802 | Improved |
| raise OPD threshold T from 2 to 4 (ablation) | OPD · contamination detection | F1: precision 0.764 -> 0.921; F1 0.802 -> 0.714 | Mixed |
| prune OPD-flagged instances from benchmark | 17 LLMs · HellaSwag | accuracy: average -22.77% | Hurt |
Closest papers by meaning
- Large Language Models Are Not Robust Multiple Choice Selectors (ICLR 2024 · similarity 0.91)
- Evaluating language models as risk scores (NeurIPS 2024 · similarity 0.90)
- B-score: Detecting biases in large language models using response history (ICML 2025 · similarity 0.90)
- A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering (NeurIPS 2026 · similarity 0.90)
- Unbiased Evaluation of Large Language Models from a Causal Perspective (ICML 2025 · similarity 0.88)
- Rethinking Generative Large Language Model Evaluation for Semantic Comprehension (ICML 2024 · 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.
