PINTO: Faithful Language Reasoning Using Prompt-Generated Rationales

TL;DR

PINTO is a prompt-based reasoning pipeline for language models that improves faithfulness and generalization by learning to reason over generated rationales with counterfactual regularization.

Contribution

It introduces a novel prompt-based learning approach with counterfactual regularization to produce more faithful and generalizable rationales in language reasoning tasks.

Findings

PINTO outperforms baselines on multiple datasets.

Rationales generated by PINTO are more faithful to model decisions.

PINTO enhances both in-distribution and out-of-distribution performance.

Abstract

Neural language models (LMs) have achieved impressive results on various language-based reasoning tasks by utilizing latent knowledge encoded in their own pretrained parameters. To make this reasoning process more explicit, recent works retrieve a rationalizing LM's internal knowledge by training or prompting it to generate free-text rationales, which can be used to guide task predictions made by either the same LM or a separate reasoning LM. However, rationalizing LMs require expensive rationale annotation and/or computation, without any assurance that their generated rationales improve LM task performance or faithfully reflect LM decision-making. In this paper, we propose PINTO, an LM pipeline that rationalizes via prompt-based learning, and learns to faithfully reason over rationales via counterfactual regularization. First, PINTO maps out a suitable reasoning process for the task input by prompting a frozen rationalizing LM to generate a free-text rationale. Second, PINTO's reasoning LM is fine-tuned to solve the task using the generated rationale as context, while regularized to output less confident predictions when the rationale is perturbed. Across four datasets, we show that PINTO significantly improves the generalization ability of the reasoning LM, yielding higher performance on both in-distribution and out-of-distribution test sets. Also, we find that PINTO's rationales are more faithful to its task predictions than those generated by competitive baselines.