RECKONING: Reasoning through Dynamic Knowledge Encoding: Generalization to Real-World knowledge

Table of Links

Abstract and 1. Introduction

2. Background

3. Method

4. Experiments

   4.1 Multi-hop Reasoning Performance

   4.2 Reasoning with Distractors

   4.3 Generalization to Real-World knowledge

   4.4 Run-time Analysis

   4.5 Memorizing Knowledge

5. Related Work

6. Conclusion, Acknowledgements, and References

A. Dataset

B. In-context Reasoning with Distractors

C. Implementation Details

D. Adaptive Learning Rate

E. Experiments with Large Language Models

4.3 Generalization to Real-World knowledge

To investigate how generalizable our method is to real-world knowledge beyond the synthetic setting, we evaluate RECKONING on a more real-world multi-hop logical reasoning task, FOLIO [29], and report the result in Table 2. The dataset has a rich vocabulary, diverse logic patterns, and abundant language variations. It has been shown to challenge LLMs in both supervised fine-tuning and in-context learning settings. We fine-tune the GPT-2 model following the in-context reasoning setting as the baseline. As before, we train the GPT-2 model and RECKONING using the multi-task objective. We also compare to more advanced baselines, including GPT-3.5 (text-davinci-003 [55]) and ChatGPT(gpt-3.5-turbo[2]), two popular large language models with around 175B parameters. For these two large models, we evaluate both in the zero-shot and few-shot settings. In the few-shot setting, we prompt the model with 8 single-task examples randomly sampled from the training set to perform in-context learning. We find that RECKONING’s performance (which is initiated here from GPT-2) is better than the GPT-2 in-context reasoning baseline. Compared to the two advanced large language models, RECKONING outperforms them by a significant margin (12% 0-shot and 7% 8-shot). We conclude that RECKONING is effective and significantly benefits reasoning tasks using real-world knowledge.

2 https://openai.com/blog/chatgpt