Overview
The surveyed methods are practical: start with CoT and retrieval for quick gains, and use fine-tuning or RLHF when you need consistent, production-grade reasoning.
Citations8
Evidence Strength0.60
Confidence0.80
Risk Signals11
Trust Signals
Findings with numeric evidence: 0/6
Findings with evidence refs: 6/6
Results with explicit delta: 0/3
Reproducibility
Status: No open assets linked
Open source: Unknown
At A Glance
Cost impact: 50%
Production readiness: 60%
Novelty: 40%
Why It Matters For Business
Better reasoning reduces wrong conclusions, lowers downstream verification cost, and enables LLMs to be used in higher-stakes workflows like finance, legal, and scientific support.
Who Should Care
Summary TLDR
This short survey groups recent techniques that improve LLM reasoning into three buckets: prompting tricks (Chain-of-Thought, Self-Consistency, Tree-of-Thought, PAL), architectural changes (retrieval augmentation, neuro-symbolic hybrids, memory modules, GNNs), and learning methods (fine-tuning on reasoning datasets, RLHF, self-supervision). It reviews common benchmarks (GSM8K, MATH, ARC, HotpotQA, LogiQA) and highlights open problems: hallucinations, adversarial fragility, domain transfer, and evaluation gaps. The paper is a practical map for engineers deciding which approach to try depending on task type and compute trade-offs.
Problem Statement
Large language models write fluent text but often fail on multi-step, logical, or formal reasoning. They hallucinate facts, produce inconsistent intermediate steps, and struggle to transfer reasoning across domains. The paper asks: which prompting, model, or training techniques reliably improve reasoning and what are their trade-offs?
Main Contribution
Organizes recent methods to boost LLM reasoning into prompting, architecture, and learning categories.
Summarizes popular benchmarks and metrics used to measure reasoning ability.
Key Findings
Chain-of-Thought prompting helps multi-step problems by making the model emit intermediate steps.
Self-Consistency improves CoT results by aggregating multiple reasoning paths and voting for the final answer.
Results
| Metric | Value | Baseline | Delta | Split / Dataset | Evidence | Evidence Ref |
|---|---|---|---|---|---|---|
| CoT vs standard prompting | Reported accuracy improvements on arithmetic and logic tasks across studies | Standard prompting | — | GSM8K, math/logic benchmarks | CoT increases multi-step task accuracy relative to vanilla prompting | [11] |
| Self-Consistency effect | Higher reliability by majority voting across multiple CoT samples | Single CoT chain | — | Math and reasoning tasks (reported across studies) | Self-Consistency aggregates multiple chains to reduce errors | [12] |
What To Try In 7 Days
Implement Chain-of-Thought prompts for multi-step QA and measure accuracy uplift on a held-out set
Add Self-Consistency: generate multiple reasoning chains and vote to stabilize answers
Prototype a retrieval step (dense or BM25) to ground answers for domain-specific queries
Reproducibility
Risks & Boundaries
Limitations
Survey summarizes prior work; it does not present new experiments or unified benchmarks.
Many claims rely on heterogeneous studies and lack standardized numeric comparisons.
When Not To Use
Do not rely on CoT alone for high-stakes decisions without external verification.
Avoid Tree-of-Thought or massive self-consistency sampling when latency or cost are tight.
Failure Modes
Hallucinated intermediate steps that look plausible but are wrong.
Error propagation along CoT chains leading to incorrect final answers.