Reproduce “Prompt-Based Bias Calibration for Better Zero/Few-Shot Learning of Language Models” on OpenResearch

About this paper

Kang He, Yinghan Long, Kaushik Roy

A new intrinsic bias calibration method for language models, named IntrCal, leverages automatically generated "null-meaning" prompts to update only a small fraction of the model's parameters, achieving an average absolute gain of up to 9% in zero-shot in-context learning and up to 7.4% in few-shot settings across various classification tasks while preserving core language modeling abilities.

Key insights

• Directly calibrating intrinsic biases within the LM's parameters provides more robust and lasting performance benefits compared to merely adjusting output probabilities.
• Updating only a minimal subset of parameters (bias parameters) is sufficient for effective bias calibration, significantly enhancing efficiency while preserving the model's core language modeling knowledge.
• Providing 'null-meaning' inputs to prompt the model into a uniform prediction state effectively establishes a 'fair starting point' for improved downstream task performance.

Summary by alphaXiv.