Measuring and Narrowing the Compositionality Gap in Language Models

By Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, Michael Lewis. 2022

1,329 citations recorded by the source corpus.

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.

Subfields: Large Language Models, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Large Language Models

In Natural Language Processing

Citations in the corpus

Self-Refine: Iterative Refinement with Self-Feedback (2023, background): “Self-ask (Press et al., 2022) ask followup ques when interim ans correct;final wrong few shot none (none)”
Augmented Language Models: a Survey (2023, background): “Press et al. (2022) call this discrepancy the compositionality gap for LMs.”
PAL: Program-aided Language Models (2022, background): “, 2022) and self-ask (Press et al., 2022), which decomposes a complicated question into a set of sub-questions.”
GPQA: A Graduate-Level Google-Proof Q&A Benchmark (2023, positive): “To enable GPT-4 to use the internet, we build on self-ask (Press et al., 2022), which provides a structured framework for multi-hop Q&A for language models with access to internet search tools.”
Enabling Large Language Models to Generate Text with Citations (2023, methodology): “In I NLI - NE S EARCH , we allow LLMs to call “search” during the generation process (Yao et al., 2023; Press et al., 2022; Jiang et al., 2023).”
The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A" (2023, methodology): “Previous research has found similar inconsistencies in LLMs (Fluri et al., 2023; Elazar et al., 2021; Press et al., 2023; Hosseini et al., 2021; Lin et al., 2022; Shi et al., 2023) Forward vs backward recall in humans Does the Reversal Curse apply to humans?”
A Comprehensive Evaluation of Tool-Assisted Generation Strategies (2023, methodology): “In this work, we use Google Search as the tool as in the original work by (Press et al., 2023).”
A Survey on In-context Learning (2022, background): “In addition to efficiency and flexibility, ICL also shows potential in safety (Panda et al., 2023), (Meade et al., 2023) use ICL for retrieved demonstrations to steer a model towards safer generations, reducing bias and toxicity in the model.”
How Language Model Hallucinations Can Snowball (2023, background): “For instance, allowing an LM to generate intermediate steps (Nye et al., 2021; Wei et al., 2022; Press et al., 2022) enables it to reach a different answer than it otherwise would.”
DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines (2023, background): “Zhang et al. (2022) reports 48% for text-davinci-002, close to our results using llama2-13b, and 63% with codex using CoT.”
Reasoning with Language Model Prompting: A Survey (2022, neutral): “, 2022b), Self-Ask (Press et al., 2022), Auto-CoT (Zhang et al.”
Language Models can Solve Computer Tasks (2023, background): “This discovery has led to the emergence of a new line of research where reasoning capabilities are used to address tasks beyond reasoning [61, 22], or enhance reasoning proficiency [26, 28, 54, 29, 62, 40, 58, 51, 12].”
Large Language Models for Information Retrieval: A Survey (2023, neutral): “To allow LLMs to actively use search engines, Self-Ask [188] and DSP [189] try to employ few-shot prompts for LLMs, triggering them to search queries when they believe it is required.”
Group-in-Group Policy Optimization for LLM Agent Training (2025, neutral): “5-1.5B-Instruct as the policy of the agent.”
Large Language Models Can Be Easily Distracted by Irrelevant Context (2023, background): “…al., 2022; Shi et al., 2022b, inter alia ), problem decomposition (Zhou et al., 2022; Drozdov et al., 2022; Dohan et al., 2022; Khot et al., 2022; Press et al., 2022, inter alia ), generating programs (Austin et al., 2021; Chowdhery et al., 2022; Gao et al., 2022; Chen et al., 2022, inter alia…”
ToolRL: Reward is All Tool Learning Needs (2025, positive): “We also evaluate on a representative QA benchmark Bamboogle (Press et al., 2022), which comprises a variety of question-answering tasks where performance is measured based on the final answer accuracy rather than the correctness of tool use.”
Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models (2024, methodology): “Figures 3 and 4 show results on HotPotQA and Bamboogle.”