Paper: Accuracy and Order Sensitivity Diverge Under Label-Free Strategies
qub · hf · 2026-08-19
This study investigates the effect of preventing models from seeing option labels during multiple-choice QA.
Findings suggest that this does not reliably reduce positional bias or improve accuracy. Baseline performance is preserved only when an LLM matcher shows all options.
More from Research
- Qwen 3.8 Comparison: Smaller Q4 Model Outreasons Larger Q5 — k-r-a-u-s-f-a-d-r · 2026-08-19
- Open Source: Acoustic UAV Detection for Battlefield Scenarios — yehors · 2026-08-19
- ClawGym II paper: Improving agents via mixed-harness training — omarsar0 · 2026-08-19
- Rich Sutton: Synthetic Data Is a Mistake, LLMs Are Only a Quarter of Intelligence — GregCook2011 · 2026-08-19
- Study: Context compactor hides real costs as Agent retrieval calls triple — dair_ai · 2026-08-19
- Paper: LLMs severely underestimate missing information, study finds — marinkazitnik · 2026-08-19