Source: Calibrate Before Use: Improving Few-Shot Performance of Language Models
Zhao et al. (2021): few-shot accuracy swings with example choice, order and format because of recency and common-token biases; contextual calibration corrects it.
Contents
- What: research paper (ICML 2021), arXiv:2102.09690
- Authors: Tony Z. Zhao, Eric Wallace, Shi Feng, Dan Klein, Sameer Singh
- Date: v1 2021-02-19, v2 2021-06-10
- Original: arXiv:2102.09690
Summary
Documents how unstable GPT-3 few-shot prompting is, traces the instability to systematic output biases, and proposes contextual calibration to correct them.
Key claims
- Accuracy depends on prompt format, which examples are chosen and even their order, ranging from near chance to near state of the art.
- Causes: recency bias (favouring answers near the end of the prompt) and common-token bias (favouring answers frequent in pre-training).
- Contextual calibration: run the prompt with a content-free input such as “N/A”, estimate the bias, then rescale so that input’s predictions are uniform.
- Calibration raised GPT-2/GPT-3 average accuracy by up to 30.0 points absolute and reduced variance across prompts.