Logically Consistent Adversarial Attacks for Soft Theorem Provers

TL;DR

This paper introduces LAVA, a novel framework combining structured generative models and symbolic solvers to create logically consistent adversarial attacks on soft theorem provers, revealing their weaknesses and improving their reasoning through training.

Contribution

The paper presents LAVA, a new method for generating logically consistent adversarial examples that expose and help improve soft theorem provers' reasoning abilities.

Findings

LAVA successfully generates adversarial attacks that reveal model weaknesses.

Training on adversarial samples improves model performance.

Identifies common logical reasoning vulnerabilities in models.

Abstract

Recent efforts within the AI community have yielded impressive results towards "soft theorem proving" over natural language sentences using language models. We propose a novel, generative adversarial framework for probing and improving these models' reasoning capabilities. Adversarial attacks in this domain suffer from the logical inconsistency problem, whereby perturbations to the input may alter the label. Our Logically consistent AdVersarial Attacker, LAVA, addresses this by combining a structured generative process with a symbolic solver, guaranteeing logical consistency. Our framework successfully generates adversarial attacks and identifies global weaknesses common across multiple target models. Our analyses reveal naive heuristics and vulnerabilities in these models' reasoning capabilities, exposing an incomplete grasp of logical deduction under logic programs. Finally, in addition to effective probing of these models, we show that training on the generated samples improves the target model's performance.