# Leap-Of-Thought: Teaching Pre-Trained Models to Systematically Reason Over Implicit Knowledge

arXiv: https://arxiv.org/abs/2006.06609 · HTML: https://mlsanity.com/paper/2006.06609/

## Key facts

- Cited by 24 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 11 Jun 2020.

## Built on by (most cited first, 24 of 24)

- [CommonsenseQA 2.0: Exposing the Limits of AI through Gamification](https://mlsanity.com/paper/2201.05320/) (2022)
- [When to Make Exceptions: Exploring Language Models as Accounts of Human Moral Judgment](https://mlsanity.com/paper/2210.01478/) (2022)
- [When Can Models Learn From Explanations? A Formal Framework for Understanding the Roles of Explanation Data](https://mlsanity.com/paper/2102.02201/) (2021)
- [Bridging the Gap: A Survey on Integrating (Human) Feedback for Natural Language Generation](https://mlsanity.com/paper/2305.00955/) (2023)
- [Datasheet for the Pile](https://mlsanity.com/paper/2201.07311/) (2022)
- [Towards Teachable Reasoning Systems: Using a Dynamic Memory of User Feedback for Continual System Improvement](https://mlsanity.com/paper/2204.13074/) (2022)
- [Learning to Repair: Repairing model output errors after deployment using a dynamic memory of feedback](https://mlsanity.com/paper/2112.09737/) (2021)
- [Chain of Thought Imitation with Procedure Cloning](https://mlsanity.com/paper/2205.10816/) (2022)
- [RuleBert: Teaching Soft Rules to Pre-trained Language Models](https://mlsanity.com/paper/2109.13006/) (2021)
- [GNN is a Counter? Revisiting GNN for Question Answering](https://mlsanity.com/paper/2110.03192/) (2021)
- [METGEN: A Module-Based Entailment Tree Generation Framework for Answer Explanation](https://mlsanity.com/paper/2205.02593/) (2022)
- [KI-BERT: Infusing Knowledge Context for Better Language and Domain Understanding](https://mlsanity.com/paper/2104.08145/) (2021)
- [Interscript: A dataset for interactive learning of scripts through error feedback](https://mlsanity.com/paper/2112.07867/) (2021)
- [Does Pre-training Induce Systematic Inference? How Masked Language Models Acquire Commonsense Knowledge](https://mlsanity.com/paper/2112.08583/) (2021)
- [Can Transformers Reason in Fragments of Natural Language?](https://mlsanity.com/paper/2211.05417/) (2022)
- [Improving Neural Model Performance through Natural Language Feedback on Their Explanations](https://mlsanity.com/paper/2104.08765/) (2021)
- [Logical Reasoning for Task Oriented Dialogue Systems](https://mlsanity.com/paper/2202.04161/) (2022)
- [Decker: Double Check with Heterogeneous Knowledge for Commonsense Fact Verification](https://mlsanity.com/paper/2305.05921/) (2023)
- [Updater-Extractor Architecture for Inductive World State Representations](https://mlsanity.com/paper/2104.05500/) (2021)
- [On Semantic Cognition, Inductive Generalization, and Language Models](https://mlsanity.com/paper/2111.02603/) (2021)
- [Logically Consistent Adversarial Attacks for Soft Theorem Provers](https://mlsanity.com/paper/2205.00047/) (2022)
- [Using Language Models For Knowledge Acquisition in Natural Language Reasoning Problems](https://mlsanity.com/paper/2304.01771/) (2023)
- [Overcoming Poor Word Embeddings with Word Definitions](https://mlsanity.com/paper/2103.03842/) (2021)
- [Concepts, Properties and an Approach for Compositional Generalization](https://mlsanity.com/paper/2102.04225/) (2021)

## Abstract (arXiv)

To what extent can a neural network systematically reason over symbolic facts? Evidence suggests that large pre-trained language models (LMs) acquire some reasoning capacity, but this ability is difficult to control. Recently, it has been shown that Transformer-based models succeed in consistent reasoning over explicit symbolic facts, under a "closed-world" assumption. However, in an open-domain setup, it is desirable to tap into the vast reservoir of implicit knowledge already encoded in the parameters of pre-trained LMs. In this work, we provide a first demonstration that LMs can be trained to reliably perform systematic reasoning combining both implicit, pre-trained knowledge and explicit natural language statements. To do this, we describe a procedure for automatically generating datasets that teach a model new reasoning skills, and demonstrate that models learn to effectively perform inference which involves implicit taxonomic and world knowledge, chaining and counting. Finally, we show that "teaching" models to reason generalizes beyond the training distribution: they successfully compose the usage of multiple reasoning skills in single examples. Our work paves a path towards open-domain systems that constantly improve by interacting with users who can instantly correct a model by adding simple natural language statements.