Source-linked AI summary
Navigate through Enigmatic Labyrinth A Survey of Chain of Thought Reasoning: Advances, Frontiers and Future
Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, Ting Liu
TL;DR
Chain-of-thought prompting enhances LLM reasoning, motivating a systematic examination of this emerging capability. The paper surveys advanced methods through a taxonomy, reviews benchmarks and frontiers, and discusses challenges, future directions, and open questions.
Problem
Chain-of-thought reasoning has attracted attention because it significantly boosts LLM reasoning abilities, especially on complex tasks.
Method
The paper systematically surveys CoT reasoning, categorizes advanced methods, and reviews benchmarks, frontiers, challenges, future directions, and open questions.
Results
The survey provides a comprehensive review of generalized CoT reasoning and organizes advanced methods through a meticulous taxonomy.
Takeaways & Limitations
The survey aims to introduce CoT reasoning to beginners and facilitate further research in the area.
Takeaways & Limitations
Due to page limitations, the survey gives only brief summaries rather than exhaustive technical details and may omit important work from venues outside its primary sources.
Abstract
from arXiv · showhide
Reasoning, a fundamental cognitive process integral to human intelligence, has garnered substantial interest within artificial intelligence. Notably, recent studies have revealed that chain-of-thought prompting significantly enhances LLM's reasoning capabilities, which attracts widespread attention from both academics and industry. In this paper, we systematically investigate relevant research, summarizing advanced methods through a meticulous taxonomy that offers novel perspectives. Moreover, we delve into the current frontiers and delineate the challenges and future directions, thereby shedding light on future research. Furthermore, we engage in a discussion about open questions. We hope this paper serves as an introduction for beginners and fosters future research. Resources have been made publicly available at https://github.com/zchuz/CoT-Reasoning-Survey
1 Introduction
Reasoning is central to human cognition, and chain-of-thought prompting enables LLMs to reason step by step, particularly on complex tasks. The paper surveys this growing research area through a taxonomy and discussion of benchmarks, frontiers, challenges, and future directions.
- Chain-of-thought reasoning enables LLMs to perform step-by-step reasoning through in-context learning, especially on complex tasks.
- CoT decomposes intricate problems into manageable thoughts and links them into a reasoning trajectory.The trajectory can make the model’s decision-making process observable.
- The paper addresses the lack of a systematic review by analyzing generalized chain-of-thought reasoning, called XoT.XoT describes gradually unraveling complex problems through step-by-step reasoning.
- The survey introduces a meticulous taxonomy, discusses frontier research, outlines challenges and future directions, and provides public resources.Its organization covers background, benchmarks, advanced methods, frontiers, future directions, and open questions.
2 Background and Preliminary
The paper places chain-of-thought reasoning within the broader shift toward prompting-based language-model use and formalizes standard and CoT prompting. CoT generates reasoning trajectories before answers, providing reasoning, interpretability, and collaboration advantages.
- As language-model pre-training scales, in-context learning and chain-of-thought reasoning emerge alongside a shift from fine-tuning toward prompting.
- Preliminary: Standard few-shot prompting supplies instructions and question-answer demonstrations, after which the model directly predicts an answer.
- Preliminary: Few-shot CoT prompting adds rationales to demonstrations and generates a step-by-step reasoning trajectory before producing the answer.
- Advantages of CoT Reasoning: CoT breaks complex problems into connected steps, exposes reasoning traces, and supports user interaction with the model’s execution trajectory.
3 Benchmarks
The survey reviews benchmarks spanning mathematical, commonsense, symbolic, logical, and multimodal reasoning. It also organizes advanced XoT methods by prompt construction and reasoning topology, as represented in its taxonomy.
- Reasoning benchmarks cover mathematical, commonsense, symbolic, logical, and multimodal capabilities.The survey provides an overview in Table 1 and directs readers to Appendix B for further benchmark details.
- Mathematical reasoning benchmarks commonly assess general LLM reasoning ability, while commonsense benchmarks assess world comprehension.
- The XoT taxonomy groups advanced methods by prompt construction and topological variation, including manual, automatic, semi-automatic, chain, and tree structures.The broader taxonomy also includes enhancement methods and is shown in Figure 2.
- Symbolic reasoning separates semantics to test whether language models can simulate atomic operations, whereas logical reasoning concerns rational thinking and robust problem-solving.
- Multimodal reasoning combines textual thought with sensory information such as visual scenes and auditory sounds.
4 Advanced Methods
The survey organizes advanced XoT methods by prompt construction, reasoning topology, and enhancement strategies. These methods trade human effort, reasoning stability, exploration, error correction, and inference cost in different ways.
- Overview: The taxonomy groups advanced XoT methods into prompt construction, topological variations, and enhancement methods.Prompt construction is further divided into manual, automatic, and semi-automatic approaches.
- Prompt Construction: Automatic and semi-automatic approaches improve demonstration selection and rationale quality through clustering, question patterns, entropy, sampling, bootstrapping, and iterative example expansion.Instruction-based zero-shot methods are described as unstable because they lack clearly defined demonstrations.
- Prompt Construction: Manual prompting offers better performance but high labor costs and weak domain transfer, while automatic prompting reduces labor but suffers errors and instability.Semi-automatic prompting balances performance and cost by incorporating limited human supervision.
- Topological Variants: Chain, tree, and graph topologies expand reasoning from sequential descriptions toward exploration, backtracking, subproblem aggregation, and self-verification.Programming, formal logic, and algorithmic descriptions can decouple thought generation from execution and provide higher-level reasoning frameworks.
- Topological Variants: Complex topological structures provide fine control flow for harder problems but limit application to general reasoning.The survey identifies this complexity as a challenge for future research.
- Enhancement Methods: Enhancement methods include verification and refinement, question decomposition, knowledge enhancement, self-ensemble, and efficient reasoning.Verification and refinement target faithful errors, while question decomposition progressively solves simpler subproblems.
- Enhancement Methods: Self-ensemble can improve performance by combining multiple reasoning chains, but its increased inference cost limits wide application.Methods include verifier ranking, self-ranking, majority voting, and complexity-based voting.
5 Frontiers of Research
The survey describes frontiers that extend XoT beyond direct text reasoning toward tools, planning, autonomous agents, and smaller deployable models. These directions improve interaction or efficiency while exposing generalization and capability trade-offs.
- Tool Use: External tools help LLMs access news, perform calculations, and interact with multimodal environments, but multiple invocations and query-error correction remain difficult.XoT supports elicitation, tracking, and action refinement alongside tool interaction.
- Planning: Planning methods decompose complex goals into sub-tasks and track execution using code, definition languages, or dynamic search algorithms.Examples include Python plans, PDDL, DFS, BFS, Monte Carlo Tree Search, and heuristic A* search.
- Autonomous Agents: The integration of planning, reasoning, memory, and tool utilization forms a cornerstone of LLM-powered autonomous agents.The survey also states that robust reasoning enables strategies for achieving complex goals.
- Distillation: Distillation transfers reasoning chains from larger models to smaller models, including natural-language, code, tabular, and preference-optimized reasoning.Self-consistency and self-distillation can generate or filter reasoning chains without external supervision.
- Distillation: Task-specific reasoning distillation may adversely downgrade general performance because language models have multidimensional capability trade-offs.This is identified as a scope boundary of the shared distillation paradigm.
6 Future Directions
The survey identifies future directions involving multimodal reasoning, faithful reasoning, and theoretical understanding of CoT. It frames these areas around integrating modalities, detecting and correcting unfaithful reasoning, and explaining how CoT emerges and helps.
- Multimodal Reasoning: Current XoT research mostly focuses on plain text, motivating multimodal research in vision-text interaction, VLLMs, and video reasoning.The survey highlights visual-textual integration, applying LLM reasoning techniques to multimodal domains, and temporal dependencies in video.
- Faithful Reasoning: LLMs often produce unfaithful reasoning through factual errors and inconsistent reasoning.Retrieval augmentation, deductive logic, post-processing, critic-based methods, neural-symbolic reasoning, and question decomposition are surveyed as responses.
- Faithful Reasoning: Faithful reasoning requires solving both detection of unfaithful reasoning and correction through accurate feedback and refinement.The survey presents these as two significant challenges.
- CoT Theory: The mechanisms behind CoT and in-context learning remain insufficiently explained despite empirical and theoretical studies.Existing work investigates why CoT enhances reasoning, but the survey characterizes current theory exploration as limited to the surface level.
7 Discussion
The discussion focuses on how CoT abilities emerge with large-scale pre-training, how to provide accurate reasoning feedback, and what CoT implies for autonomous agents and AGI.
- Open Questions: The open-question discussion examines CoT emergence, accurate feedback for reasoning and decision-making, and implications for autonomous agents and AGI.These topics are discussed in Appendix A.2.
8 Conclusion
The paper systematically surveys generalized chain-of-thought reasoning, categorizing advanced methods, examining frontiers and challenges, and identifying future directions and open questions.
- The survey provides a comprehensive review of existing research on generalized chain-of-thought reasoning.
- It meticulously categorizes advanced methods and examines current frontier research.
- The paper highlights existing challenges and identifies potential future research directions.
- It also discusses open questions to facilitate further research in the area.
Limitations
The survey distinguishes its generalized XoT focus from prior specialized or broader surveys, while acknowledging incomplete technical coverage, venue omissions, and limited benchmark completeness.
- This study presents the first comprehensive survey of generalized chain-of-thought reasoning.
- Page limitations restrict each method to a brief summary rather than exhaustive technical details.
- The survey mainly collects studies from selected venues and arXiv, so important work from other venues may be missing.
- Its benchmarks section primarily lists widely used datasets, while more complete benchmarks are available elsewhere.
- Unlike prior surveys, it focuses on generalized chain-of-thought reasoning in the era of large language models.
A.2 Further Discussion
The discussion connects chain-of-thought reasoning to code pre-training, feedback generation, autonomous agents, and broader AGI debates, while identifying unresolved questions.
- The source of chain-of-thought ability in code-data pre-training remains an open question despite several indications supporting that possibility.
- Mixing code data during pre-training enhances general reasoning, whereas instruction fine-tuning with code data provides task-specific reasoning abilities.
- Current feedback methods include model self-feedback, other-model feedback, external-environment feedback, and reinforcement-learning-based feedback.
- No fully satisfying approach currently exists for accurately obtaining feedback signals from intermediate reasoning.
- LLM-powered AI is progressing toward more generalized behavior, while whether LLMs can evolve into AGI remains intensely debated.
A.4 Empirical Results
The empirical-results section surveys XoT performance across mathematical, commonsense, symbolic, logical, and multimodal reasoning, presenting broad trends rather than strictly comparable estimates.
- The survey compiles XoT performance in mathematics, commonsense, and symbolic reasoning, primarily using GPT-series models.
- Because model checkpoints and experimental setups vary, results using the same backbone LLM may not be fairly comparable.
- The reported table therefore provides only a rough trend of performance.
- Commonsense reasoning benchmarks evaluate inference, judgment, and insight based on widely known world knowledge.
- Common symbolic reasoning tasks include last-letter concatenation, coin flip, reverse list, state tracking, and object counting.
- Logical reasoning is organized into deductive, inductive, and abductive forms.
- Visual and video multimodal reasoning benchmarks extend evaluation beyond text, with ScienceQA noted as a commonly used visual benchmark.
B.6 Comprehensive Benchmarks
The survey organizes reasoning evaluation around comprehensive benchmarks and presents empirical results for XoT methods across several reasoning settings. It also situates these benchmarks within a broader taxonomy of methods, frontiers, and future directions.
- Empirical Evaluation: Table 2 reports various XoT methods on mathematical, commonsense, and symbolic reasoning benchmarks.Listed methods include few-shot and zero-shot CoT, PAL, PoT, Plan-and-Solve, Auto-CoT, and other prompt-construction approaches.
- Evaluation Caveat: Performances in Table 2 are not directly comparable because the evaluated methods use different experimental setups.The table is intended to provide an overall empirical insight rather than a controlled ranking of methods.
- Survey Organization: Figure 8 presents a full-edition overview connecting advanced methods, frontiers, future directions, and benchmarks.The figure frames benchmark coverage as part of the survey’s broader organization of XoT reasoning research.