Irrelevant Context Alters Model Answers
sanmikoyejo · x · 2026-07-16
This study tests how models perform when task-irrelevant context is prepended to questions across multiple QA benchmarks.
Key findings:
- Even semantically meaningless pseudo-words can cause models to significantly change their answers on certain samples
- This effect leads to both "worse" and "better" outcomes for different samples
- The upper and lower tail samples (those with the most significant changes) account for the majority of the overall variance
This highlights the models' sensitivity to contextual noise, suggesting that we must look at distribution tails rather than just average scores.
More from Research
- HarnessMM cuts agent tokens per task by up to 88% while raising success rates — omarsar0 · 2026-10-26
- Ben Recht: The world needs a truly open language model, and it's easier to build than you think — beenwrekt · 2026-10-26
- New Blog Breaks Black-Box Distillation Into Five Practical Steps — chrisrohlf · 2026-10-26
- DetectorScope turns black-box AI text detectors into auditable ones with sparse autoencoders — _akhaliq · 2026-10-26
- OPST paper shows LLMs self-improve with zero human data, guided only by topic descriptions — _akhaliq · 2026-10-26
- Tencent's HiSTEVE 1.0: hierarchical VLA agent plays Minecraft with native keyboard and mouse, hits 85% SOTA — _akhaliq · 2026-10-26