Varying Shades of Wrong: Aligning LLMs with Wrong Answers Only

TL;DR

This paper explores how large language models can be aligned to prefer certain wrong answers over others, revealing that such alignment can reduce errors and improve calibration even without reliable annotations.

Contribution

It introduces methods to elicit and utilize wrong-over-wrong preferences for aligning LLMs, demonstrating their effectiveness across multiple models and datasets.

Findings

LLMs can distinguish shades of wrong answers with up to 20.9% better than random.

Aligning with wrong-over-wrong preferences reduces errors and sometimes yields correct answers.

Alignment improves model calibration and understanding of wrong options.

Abstract

In the absence of abundant reliable annotations for challenging tasks and contexts, how can we expand the frontier of LLM capabilities with potentially wrong answers? We focus on two research questions: (1) Can LLMs generate reliable preferences among wrong options? And if so, (2) Would alignment with such wrong-over-wrong preferences be helpful? We employ methods based on self-consistency, token probabilities, and LLM-as-a-judge to elicit wrong-over-wrong preferences, and fine-tune language models with preference optimization approaches using these synthesized preferences. Extensive experiments with seven LLMs and eight datasets demonstrate that (1) LLMs do have preliminary capability in distinguishing various shades of wrong, achieving up to 20.9% higher performance than random guess; (2) Alignment with wrong-over-wrong preferences helps LLMs to produce less wrong and sometimes even outright correct answers, while overall improving model calibration.