Source-linked AI summary
Reasoning with Language Model Prompting: A Survey
Shuofei Qiao, Yixin Ou, Ningyu Zhang, Xiang Chen, Yunzhi Yao, Shumin Deng, Chuanqi Tan, Fei Huang, Huajun Chen
TL;DR
Reasoning is important for complex problem-solving, yet language models have required prompting and scaling to exhibit broader reasoning abilities. This paper surveys reasoning with language-model prompting, compares and categorizes existing methods, and provides resources and future directions. It concludes with a review of current methods and a vision for more effective prompting models, while acknowledging coverage and categorization limitations.
Problem
Reasoning supports complex problem-solving, but natural-language-processing neural networks have struggled to reason from stated or previously learned information.
Method
The paper surveys reasoning methods with language-model prompting, categorizing them into strategy-enhanced and knowledge-enhanced reasoning and comparing representative approaches.
Results
The survey provides comprehensive comparisons, research summaries, systematic resources, discussion of possible emerging reasoning abilities, and future research directions.
Takeaways & Limitations
The paper envisions synergy between NLP and other domains and increasingly sophisticated, efficient language-model prompting models for reasoning performance.
Takeaways & Limitations
The survey may miss important references and technical details, and its benchmark task definitions and boundaries may not be fully accurate or exhaustive.
Abstract
from arXiv · showhide
Reasoning, as an essential ability for complex problem-solving, can provide back-end support for various real-world applications, such as medical diagnosis, negotiation, etc. This paper provides a comprehensive survey of cutting-edge research on reasoning with language model prompting. We introduce research works with comparisons and summaries and provide systematic resources to help beginners. We also discuss the potential reasons for emerging such reasoning abilities and highlight future research directions. Resources are available at https://github.com/zjunlp/Prompt4ReasoningPapers (updated periodically).
1 Introduction
Reasoning is central to complex problem-solving, but neural networks have historically struggled with it. This survey reviews how prompting and model scaling support arithmetic, commonsense, and symbolic reasoning, while organizing recent work and resources.
- Neural networks have historically struggled to reason from stated or previously learned information.
- Larger language models have exhibited arithmetic, commonsense, and symbolic reasoning abilities.
- Prompting strategies, including chain-of-thought prompting, can unlock language-model reasoning abilities.
- The survey presents recent research with comparisons, systematic organization, open resources, and future research directions.
2 Preliminaries
The survey formalizes prompted reasoning as generating answers and, for chain-of-thought methods, intermediate reasoning steps. It organizes methods into strategy enhancement and knowledge enhancement, with strategy methods further covering prompt design, process optimization, and external engines.
- Preliminaries: Standard prompting maximizes the language model likelihood of an answer A given question Q and prompt T.
- Preliminaries: Few-shot prompting defines T using K question–answer exemplars, while chain-of-thought prompting adds reasoning steps to those exemplars.
- Preliminaries: Chain-of-thought formulations factor generation into reasoning steps C followed by answer A conditioned on C.
- Strategy Enhanced Reasoning: Strategy-enhanced reasoning includes prompt engineering, process optimization, and external reasoning engines.
- Prompt Engineering: Prompt engineering distinguishes single-stage methods that improve T from multi-stage methods that append or redesign context for each reasoning step.
- Process Optimization and External Engines: Process optimization includes self-, ensemble-, and iterative-optimization methods, while external engines generate prompts, execute reasoning, or provide tool calls.
- Knowledge Enhanced Reasoning: Knowledge-enhanced reasoning uses implicit model knowledge or retrieved explicit knowledge as prompts.
3 Taxonomy of Methods
The survey organizes reasoning with language model prompting into strategy-enhanced and knowledge-enhanced methods. Strategy enhancement spans prompt engineering, process optimization, and external engines, while knowledge enhancement uses implicit or explicit knowledge.
- Taxonomy: The taxonomy divides reasoning methods into Strategy Enhanced Reasoning and Knowledge Enhanced Reasoning.The survey further refines these categories according to distinctive method features.
- Prompt Engineering: Prompt engineering distinguishes single-stage prompts from multi-stage prompts based on the number of input-output stages.Single-stage methods improve one prompt, whereas multi-stage methods add context or design prompts across reasoning stages.
- Prompt Engineering: CoT prompting adds intermediate reasoning steps to few-shot exemplars, but performance is highly sensitive to exemplar selection.Even tiny exemplar changes can cause large performance drops, making exemplar quality particularly important.
- Process Optimization: Process optimization calibrates reasoning through self-optimization, ensemble-optimization, or iterative optimization.These approaches respectively add optimizer modules, combine multiple reasoning processes, or repeatedly fine-tune the language model on generated reasoning examples.
- External Engine: External engines support reasoning as prompt producers, reasoning executors, or tool extenders.Examples include physical simulators, code interpreters, and integrated calculators, search engines, or question-answering systems.
- Knowledge Enhanced Reasoning: Knowledge-enhanced methods prompt language models with implicit knowledge generated by the model or explicit knowledge retrieved from external resources.The survey treats these as two branches of knowledge enhancement.
4 Comparison and Discussion
The survey compares reasoning methods across model scales, prompt sources, and training corpora. It finds that CoT prompting depends strongly on large models and high-quality rationales, while code pre-training may trigger stronger reasoning.
- Comparison of Language Models: Few-shot prompting improves across almost all tasks as model scale increases, while CoT gains are greatest with PaLM-540B.Below 100B parameters, CoT may provide no gain or even reduce performance.
- Comparison of Language Models: At the same parameter scale, Codex significantly outperforms GPT-3, with training corpus as their major difference.The survey associates code pre-training with reasoning benefits beyond code generation and understanding.
- Comparison of Prompts: Manual, LM-generated, and retrieval-based prompts differ in flexibility, knowledge provision, and retrieval cost.Retrieval-based prompts can reduce generation instability but often require annotated external resources and expensive information retrieval.
- Comparison of Prompts: CoT prompting works only on large LMs, whereas smaller LMs benefit from fine-tuning with rationales.The survey identifies high-quality reasoning rationales in the input context as key to reasoning with LM prompting.
- Comparison of Prompts: The reason CoT prompting succeeds remains poorly understood despite hypotheses involving code training and instruction-like rationales.The survey presents these explanations as open questions rather than established causes.
5 Benchmarks and Resources
The survey organizes reasoning benchmarks into arithmetic, commonsense, logical, symbolic, and multimodal categories. It also highlights open tools and communities that support experimentation with prompting and reasoning systems.
- Benchmarks: Arithmetic reasoning evaluates solving mathematical word problems, progressing from small single-step datasets toward more complex and larger benchmarks.It is also referred to as mathematical reasoning.
- Benchmarks: Commonsense reasoning requires LMs to incorporate physical and human interactions while assuming general background knowledge.CommonsenseQA is identified as the most widely used benchmark.
- Benchmarks: Logical reasoning includes deductive, inductive, and abductive forms, with datasets covering rule bases, conclusions, and multi-step entailment trees.Deduction proceeds from general information to specific conclusions, whereas induction generalizes from observations.
- Benchmarks: Symbolic reasoning narrowly covers simple manipulation tasks such as last-letter concatenation, reverse list, and coin flip.The survey distinguishes this task collection from symbolic AI as a broader concept.
- Benchmarks: Multimodal benchmarks extend reasoning beyond text, including ScienceQA with diverse science questions, answers, and explanations.These resources target information available across modalities such as images and other non-text inputs.
- Resources: ThoughtSource, LangChain, λprompt, and EasyInstruct provide public communities, libraries, and packages for developing and testing LM reasoning applications.Their uses include CoT resources, combining LMs with external computation or knowledge, and instructing LLMs in experiments.
6 Future Directions
The survey identifies future directions spanning theory, efficiency, robustness, multimodality, and generalization. These directions address why reasoning emerges, how to reduce resource demands, and how to extend reasoning beyond current textual and in-distribution settings.
- Theoretical Principle of Reasoning: Future theoretical work should explain emergent reasoning, including the roles of in-context learning, rationales, prompting algorithms, and model neurons.Code-pretrained LMs may handle structured reasoning better, but code-based pre-training depends on off-the-shelf structure.
- Efficient Reasoning: Efficient reasoning should reduce reliance on large LMs and account for computing resources, carbon emissions, and energy use.The survey suggests small LMs and models that generalize across evaluation scenarios as possible directions.
- Robust, Faithful and Interpretable Reasoning: Reasoning systems need greater robustness, faithfulness, and interpretability because zero-shot CoT can produce toxicity and biases.The survey also points to multi-stage selection-inference architectures as one response.
- Multimodal (Interactive) Reasoning: Multimodal reasoning should integrate images, audio, and video into unified CoT rather than merely converting images into text.Interactive reasoning among models is another proposed direction.
- Generalizable (True) Reasoning: True reasoning generalization requires solving unseen related tasks, but current out-of-distribution generalization remains far from satisfactory.Existing work has examined changes in the length of reasoning questions.
7 Conclusion and Vision
The paper reviews reasoning with language model prompting, offers comprehensive comparisons and research directions, and envisions stronger cooperation between NLP and other fields. It anticipates more sophisticated and efficient prompting models contributing to reasoning performance.
- Conclusion and Vision: The survey combines a review of reasoning with LM prompting, comprehensive comparisons, and several proposed research directions.Its vision is greater synergy between NLP and other domains alongside more sophisticated and efficient prompting models.
Limitations
The survey acknowledges limitations in its coverage, technical detail, benchmark definitions, and empirical conclusions. It plans to update the review and address gaps as the field evolves.
- Scope: The authors acknowledge that limitations remain despite their efforts and intend to update the latest opinions as the area evolves.This signals that the survey is presented as an evolving resource rather than a final account.
- References & Methods: The survey may omit important references and technical details because of page limits.Its review mainly covers cutting-edge methods from the preceding two years, mostly 2022, and will be supplemented with newer work.
- Benchmarks: Benchmark definitions and boundaries may be inaccurate because the benchmarks were gathered from experiments in mainstream works.The survey may also miss reasoning tasks such as generics, default inheritance, and non-monotonic reasoning.
- Empirical Conclusions: The survey’s conclusions and future directions are based on empirical analyses that may not be sufficiently macroscopic.The authors state that these conclusions are proposed and further speculated upon existing works.
A.1 Related Survey
Because the area is relatively nascent, only a few surveys exist. This survey distinguishes itself by systematically reviewing reasoning with language model prompting to compare methods and inspire new ideas.
- Related Surveys: Existing surveys cover adjacent areas including reasoning with large language models, in-context learning, large language models, and commonsense reasoning.The related work spans surveys of techniques, tasks, datasets, methods, and unresolved challenges.
- Distinctive Scope: This survey focuses specifically on reasoning with language model prompting rather than the broader or neighboring topics covered by prior surveys.Its stated goal is to systematically understand methodologies, compare methods, and inspire new ideas.
A.2 Taxonomy of Methods and Tasks
The survey organizes reasoning with language model prompting through a taxonomy of methods and tasks, then illustrates the taxonomy with benchmark overviews, scale comparisons, and ChatGPT cases. Its examples span arithmetic, commonsense, and symbolic reasoning.
- Taxonomy: The survey presents a complete taxonomy of language model prompting methods and reasoning tasks in Figure 8.The taxonomy is organized around methods and tasks.
- Scale Comparisons: Performance comparisons examine language models of different scales on GSM8K for arithmetic reasoning and CommonsenseQA for commonsense reasoning.The CommonsenseQA comparison extends the discussion of generalization across reasoning tasks.
- Benchmarks and Tasks: The benchmark overview summarizes reasoning datasets and tasks, with additional key statistics listed in Table 2.The survey also notes broader benchmarks for diverse and generalized reasoning capabilities.
- ChatGPT Cases: ChatGPT produces a reasoning process and correct answer for a GSM8K math problem without in-context exemplars.The reasoning process has the same format as the GSM8K gold label, which may indicate GSM8K appeared in ChatGPT’s training corpus.
- ChatGPT Cases: On CommonsenseQA, ChatGPT gives the correct answer and explains why each option is right or wrong, beyond the dataset’s gold label.The case is presented as evidence of strong commonsense reasoning ability.
- ChatGPT Cases: On Last Letter Concatenation, ChatGPT accurately describes the task but fails to answer the question, indicating weaker symbolic reasoning than in the other cases.The survey contrasts this symbolic reasoning result with its arithmetic and commonsense examples.