Table of Links
Abstract and 1. Introduction
-
Background
-
Method
-
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
-
Related Work
-
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
4.2 Reasoning with Distractors
In cases where multiple questions must be answered about the same knowledge set, some knowledge that is relevant to one question will likely be irrelevant to another question. For example, in Table 7, the fact “Charlie is White.” is not needed to answer the question “Harry is red?”. Thus, it is important to evaluate the robustness of RECKONING when there exists irrelevant information (i.e., distractors) in the knowledge set. In this experiment, we analyze RECKONING’s ability to focus on the correct knowledge and ignore distractors when answering questions. We use ProofWriter as the evaluation dataset since it already has a setting with distractors included in the knowledge. For systematic analysis, we gradually add distractors to the context (starting from 2 and finishing at all possible distractors, of which there are an average of 7 per question). We train RECKONING and the baseline using the multi-task objective, where the model must (1) recall all of the facts and rules relevant to the question and (2) predict the conclusion based on the correct knowledge. In this case, we adapt training such that for each question x, the outer-loop (Equation (5)) CLM loss is only computed with respect to the relevant facts from K, thereby learning to recall only relevant facts during training.
In Figure 5, we see that RECKONING’s performance is consistently more robust under distractors than the FT-ICR baseline. When we include all of the distractors in the context, RECKONING achieves a significantly higher average label accuracy (82.5%) across hops than the baseline (70.9%), as computed by the average of the 3 considered hop depths. Additionally, compared to performance with no distractors, RECKONING’s performance only drops 17.1% while the baseline performance drops 28.6%, thereby exhibiting a better ability to disentangle the correct knowledge from the distractors.
Finally, we also explore RECKONING’s generalizability to models with a larger parameter size. We scale up the language model we used, GPT-2-small (124M), to GPT-2-XL (1.5B) by adopting a parameter efficient finetuning method LoRA [33]. For simplicity, we only evaluate the models on the most difficult settings, i.e., ProofWriter-5-hop with all the distractors. With GPT-2-XL-LoRA, in-context reasoning achieves 65% accuracy on the test set, while our RECKONING model achieves 70.2% accuracy, a 5% performance gain. This result suggests that RECKONING’s advantages in the presence of distractors hold even as models scale in size.
:::info
Authors:
(1) Zeming Chen, EPFL (zeming.chen@epfl.ch);
(2) Gail Weiss, EPFL (antoine.bosselut@epfl.ch);
(3) Eric Mitchell, Stanford University (eric.mitchell@cs.stanford.edu)’;
(4) Asli Celikyilmaz, Meta AI Research (aslic@meta.com);
(5) Antoine Bosselut, EPFL (antoine.bosselut@epfl.ch).
:::
:::info
This paper is available on arxiv under CC BY 4.0 DEED license.
:::
