Towards Reasoning in Large Language Models: A Survey

TL;DR

Survey on reasoning in LLMs, highlighting Chain-of-Thought prompting, achieving 74.4% on GSM8K.

cs.CL 🔴 Advanced 2022-12-21 33 views
Jie Huang Kevin Chen-Chuan Chang
reasoning LLMs Chain-of-Thought evaluation future directions

Key Findings

Methodology

The paper surveys techniques to enhance reasoning in LLMs, including supervised fine-tuning, Chain-of-Thought (CoT) prompting, problem decomposition, and self-consistency decoding, validated through datasets and experiments.

Key Results

  • CoT prompting improved GSM8K accuracy from 55.7% to 74.4%, outperforming direct generation.
  • Problem decomposition excelled in compositional reasoning tasks like dynamic semantic parsing.
  • Self-consistency decoding enhanced robustness by sampling diverse reasoning paths and selecting the most consistent answer.

Significance

This study provides a systematic roadmap for reasoning in LLMs, addressing key bottlenecks in multi-step reasoning tasks and offering actionable insights for academia and industry.

Technical Contribution

Introduced novel techniques like CoT variants (e.g., Zero-shot CoT) and self-consistency decoding, demonstrating their effectiveness across diverse reasoning tasks.

Novelty

First to systematically classify and evaluate reasoning enhancement techniques, proposing innovative methods like problem decomposition and self-consistency decoding, outperforming traditional prompting.

Limitations

  • Current methods rely heavily on large-scale models, limiting applicability to smaller models.
  • Lack of unified evaluation standards for reasoning abilities.
  • Limited generalization to cross-domain tasks.

Future Work

Future work includes efficient prompting techniques, unified evaluation frameworks, and reasoning enhancement for smaller models.

AI Executive Summary

Reasoning is a cornerstone of human intelligence but remains a challenge in AI. Recent advances in large language models (LLMs) like GPT-3 and PaLM suggest emergent reasoning capabilities at scale. However, these models struggle with multi-step reasoning tasks and lack consistent evaluation frameworks.

This paper surveys techniques to enhance reasoning in LLMs, such as Chain-of-Thought (CoT) prompting, problem decomposition, and self-consistency decoding. CoT prompting explicitly guides reasoning by providing intermediate steps, significantly improving performance on tasks like GSM8K. Problem decomposition breaks complex tasks into manageable subproblems, while self-consistency decoding enhances robustness by exploring diverse reasoning paths.

Despite progress, challenges remain, including reliance on large models and inconsistent evaluation standards. Future research should focus on efficient prompting, evaluation unification, and extending reasoning to smaller models, paving the way for more robust AI systems capable of human-like reasoning.

Deep Analysis

Background

Reasoning in AI has long been a challenge. Recent LLMs like GPT-3 and PaLM exhibit emergent reasoning abilities, especially with techniques like Chain-of-Thought prompting. However, multi-step reasoning remains a bottleneck.

Core Problem

LLMs often fail in multi-step reasoning tasks, relying on heuristics rather than true logical inference. This limits their utility in complex scenarios requiring compositional reasoning.

Innovation

Key innovations include Chain-of-Thought prompting (explicit reasoning steps), problem decomposition (breaking tasks into subproblems), and self-consistency decoding (sampling diverse paths for robust answers).

Methodology

  • �� Supervised fine-tuning: Training on specific datasets with reasoning paths.
  • �� Chain-of-Thought prompting: Providing ⟨input, reasoning, output⟩ examples.
  • �� Self-consistency decoding: Sampling diverse paths and selecting the most consistent answer.
  • �� Problem decomposition: Dividing complex tasks into simpler subproblems.

Experiments

Experiments used GSM8K, MathQA, and other datasets to evaluate CoT prompting and self-consistency decoding. Ablation studies analyzed the contribution of each component.

Results

CoT prompting improved GSM8K accuracy from 55.7% to 74.4%. Problem decomposition excelled in dynamic semantic parsing. Self-consistency decoding enhanced answer consistency.

Applications

Techniques apply to math reasoning, commonsense QA, and complex semantic parsing, particularly for multi-step reasoning tasks.

Limitations & Outlook

Methods rely on large models, lack unified evaluation standards, and show limited cross-domain generalization.

Plain Language Accessible to non-experts

Imagine solving a jigsaw puzzle. Chain-of-Thought prompting is like having step-by-step instructions for assembling pieces. Problem decomposition breaks the puzzle into smaller sections. Self-consistency decoding tries multiple arrangements and picks the one that fits best.

ELI14 Explained like you're 14

Think of it like a video game! Chain-of-Thought is like a walkthrough guide showing every step. Problem decomposition is breaking a big boss fight into smaller stages. Self-consistency decoding is trying different strategies and picking the best one. Cool, right?

Glossary

Chain-of-Thought (CoT)

A prompting technique that guides models to generate intermediate reasoning steps.

Used to improve multi-step reasoning in LLMs.

Problem Decomposition

Breaking complex tasks into simpler subproblems for easier resolution.

Applied in dynamic semantic parsing tasks.

Self-Consistency Decoding

Sampling diverse reasoning paths and selecting the most consistent answer.

Enhances robustness in reasoning tasks.

Supervised Fine-tuning

Training models on specific datasets to generate reasoning paths.

Used for smaller models to improve reasoning.

GSM8K

A dataset for evaluating arithmetic reasoning with 8,000 problems.

Benchmark for testing CoT prompting.

Open Questions Unanswered questions from this research

  • 1 How to adapt reasoning techniques for smaller models?
  • 2 How to establish unified reasoning evaluation frameworks?
  • 3 How to improve cross-domain generalization?

Applications

Immediate Applications

Math Education

Helps students solve complex math problems with detailed reasoning paths.

Smart Assistants

Enhances multi-step reasoning in commonsense QA systems.

Long-term Vision

General AI

Advances reasoning capabilities toward human-like intelligence.

Abstract

Reasoning is a fundamental aspect of human intelligence that plays a crucial role in activities such as problem solving, decision making, and critical thinking. In recent years, large language models (LLMs) have made significant progress in natural language processing, and there is observation that these models may exhibit reasoning abilities when they are sufficiently large. However, it is not yet clear to what extent LLMs are capable of reasoning. This paper provides a comprehensive overview of the current state of knowledge on reasoning in LLMs, including techniques for improving and eliciting reasoning in these models, methods and benchmarks for evaluating reasoning abilities, findings and implications of previous research in this field, and suggestions on future directions. Our aim is to provide a detailed and up-to-date review of this topic and stimulate meaningful discussion and future work.

cs.CL cs.AI