Reasoning Elicitation in Language Models via Counterfactual Feedback

TL;DR

This paper introduces new metrics and fine-tuning methods to enhance the causal reasoning abilities of language models, especially in counterfactual question answering, and evaluates their effectiveness in realistic scenarios.

Contribution

It proposes novel metrics for assessing reasoning in language models and develops fine-tuning approaches to improve their causal reasoning capabilities.

Findings

Fine-tuning improves reasoning accuracy in counterfactual questions.

New metrics better capture reasoning abilities than traditional metrics.

Models show enhanced generalization in reasoning tasks after fine-tuning.

Abstract

Despite the increasing effectiveness of language models, their reasoning capabilities remain underdeveloped. In particular, causal reasoning through counterfactual question answering is lacking. This work aims to bridge this gap. We first derive novel metrics that balance accuracy in factual and counterfactual questions, capturing a more complete view of the reasoning abilities of language models than traditional factual-only based metrics. Second, we propose several fine-tuning approaches that aim to elicit better reasoning mechanisms, in the sense of the proposed metrics. Finally, we evaluate the performance of the fine-tuned language models in a variety of realistic scenarios. In particular, we investigate to what extent our fine-tuning approaches systemically achieve better generalization with respect to the base models in several problems that require, among others, inductive and deductive reasoning capabilities.