Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Measuring and Narrowing the Compositionality Gap in Language Models

About

We investigate the ability of language models to perform compositional reasoning tasks where the overall solution depends on correctly composing the answers to sub-problems. We measure how often models can correctly answer all sub-problems but not generate the overall solution, a ratio we call the compositionality gap. We evaluate this ratio by asking multi-hop questions with answers that require composing multiple facts unlikely to have been observed together during pretraining. In the GPT-3 family of models, as model size increases we show that the single-hop question answering performance improves faster than the multi-hop performance does, therefore the compositionality gap does not decrease. This surprising result suggests that while more powerful models memorize and recall more factual knowledge, they show no corresponding improvement in their ability to perform this kind of compositional reasoning. We then demonstrate how elicitive prompting (such as chain of thought) narrows the compositionality gap by reasoning explicitly. We present a new method, self-ask, that further improves on chain of thought. In our method, the model explicitly asks itself (and answers) follow-up questions before answering the initial question. We finally show that self-ask's structured prompting lets us easily plug in a search engine to answer the follow-up questions, which additionally improves accuracy.

Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, Mike Lewis• 2022

Related benchmarks

TaskDatasetResultRank
Multi-hop Question Answering2WikiMultihopQA
EM40.5
559
Multi-hop Question AnsweringHotpotQA (test)
F154.49
311
Question AnsweringOpenBookQA
Accuracy91.8
305
Multi-hop Question AnsweringHotpotQA
F1 Score64.74
294
Multitask Language UnderstandingMMLU-Pro
Accuracy41.5
248
Multi-hop Question Answering2WikiMultiHopQA (test)
EM40.1
226
Mathematical ReasoningMATH 500
Accuracy30
221
Multi-hop Question AnsweringMuSiQue
EM25
209
Multi-hop Question Answering2WikiMQA
F1 Score52.1
161
Medical Question AnsweringMedQA
Accuracy82.88
154
Showing 10 of 92 rows
...

Other info

Code

Follow for update