Teacher-Student Training for Debiasing: General Permutation Debiasing for Large Language Models

TL;DR

This paper explores two variants of student models: one based on pure distillation, and the other on an error-correction approach for more complex tasks, where the student corrects a single biased decision from the teacher to achieve a debiased output.

Abstract

Large Language Models (LLMs) have demonstrated impressive zero-shot capabilities and versatility in NLP tasks, however they sometimes fail to maintain crucial invariances for specific tasks. One example is permutation sensitivity, where LLMs' outputs may significantly vary depending on the order of the input options. While debiasing techniques can mitigate these issues, and yield better performance and reliability, they often come with a high computational cost at inference. This paper addresses this inefficiency at inference time. The aim is to distill the capabilities of a computationally intensive, debiased, teacher model into a more compact student model. We explore two variants of student models: one based on pure distillation, and the other on an error-correction approach for more complex tasks, where the student corrects a single biased decision from the teacher to achieve a debiased output. Our approach is general and can be applied to both black-box and white-box LLMs. Furthermore, we demonstrate that our compact, encoder-only student models can outperform their larger, biased teacher counterparts, achieving better results with significantly fewer parameters.

Figures (5)

• Figure 1: High-level diagram of the work: The left-hand side illustrates how LLMs may be sensitive to input ordering, but by averaging results from different permutations can yield debiased distributions. The right-hand side shows two variants of students that emulate the debiased teacher distribution, either through distillation or through error correction where the student improves a single sampled biased decision.
• Figure 2: Templates used for prompting LLMs for MCQA. For context-free questions, the context is omitted.
• Figure 3: Prompts used for comparative assessment. Different attributes use different adjectives.
• Figure 4: RACE++ performance of error correction students when using $N$ black-box samples to approximate the debiased distribution (§ \ref{['ssec:black-box']})
• Figure 5: DeBERTa-large accuracy when using a limited number of examples during teacher-student training.