Contrastive Chain-of-Thought Prompting

1 Introduction

• Prior findings showed little performance difference between valid and invalid CoT demonstrations, leaving how models learn effective reasoning unclear.Errors in intermediate reasoning can affect final-answer accuracy and model trustworthiness.Source: §1 · pp. 1–2
• Contrastive CoT provides both positive and negative reasoning demonstrations to guide step-by-step reasoning and help models avoid faults.The approach is inspired by learning from positive and negative examples.Source: §1 · p. 2
• The authors propose automatically constructing contrastive demonstrations from existing valid reasoning chains and report improvements across multiple reasoning benchmarks.They describe the method as task-agnostic and compatible with self-consistency.Source: §1 · p. 2
• GPT-3.5-Turbo gains 9.8 points on GSM-8K and 16.0 points on Bamboogle over conventional CoT.The paper also reports reduced errors in generated reasoning chains.Source: §1 · p. 2

2 Preliminary Study: Effect of Different Types of Contrastive Demonstrations

• The study tests whether including both valid and invalid rationales in demonstrations can enhance chain-of-thought prompting.The question is motivated by prior findings that invalid demonstrations had little effect despite the apparent importance of sound reasoning.Source: §2 · p. 2

2.1 Components of Chain of Thought

• CoT rationales contain bridging objects, such as numbers or entities, and language templates that guide their derivation and contextualization.Source: §2.1 · p. 2

2.2 What is Invalid Chain of Thought?

• Invalid rationales are classified by coherence and relevance, with logical mistakes included as a further category; Figure 2 presents five main categories.Coherence concerns step ordering, while relevance concerns whether the rationale corresponds to question information.Source: §2.2 · pp. 2–3, Figure 2 · p. 3

2.3 Experimental Setup

• The preliminary experiments use GPT-3.5-Turbo (0301), four-shot prompts, and GSM8K and Bamboogle, with demonstrations manually constructed by prior work.The two datasets represent arithmetic and factual reasoning, respectively.Source: §2.3 · p. 3

2.4 Preliminary Results

• All tested invalid-rationale categories improve over conventional CoT, with incoherent-object demonstrations yielding the highest average performance on GSM8K and Bamboogle.The results are reported in the preliminary-results table.Source: §2.4 · pp. 3–4, §2.2 · p. 3, Table 1 · p. 4

3 Contrastive Chain of Thought

• The approach responds to limited understanding of chain-of-thought learning, including prior findings that invalid demonstrations can yield performance similar to valid ones.The paper also notes that conventional chain of thought does not identify faults to avoid, which could contribute to error propagation.Source: §3 · pp. 4–5
• The paper introduces automatic construction to address concerns that manually created invalid rationales may not generalize well to new tasks.Prior analysis found that providing both valid and invalid demonstrations improved reasoning performance.Source: §3 · p. 4
• Contrastive chain of thought pairs correct and incorrect reasoning steps in demonstrations, while test-time prompts provide the question and demonstrations for the model to generate its own reasoning and answer.The demonstration format includes the question, reasoning steps, and answer for both the positive and negative examples.Source: §3 · p. 5
• The method automatically constructs incorrect rationales by extracting object spans from correct rationales and randomly shuffling their positions to create incoherent bridges.The extracted spans can include numbers, equations, or persons.Source: §3 · p. 5

4.1 Experimental Setup

• The evaluation covered five arithmetic datasets and two factual-QA datasets, sampling up to 500 test examples per dataset or using all examples when fewer were available.The experiments focused on arithmetic reasoning and factual question answering.Source: §4.1 · p. 5

4.2 Main Results

• Contrastive CoT consistently outperformed conventional CoT, with gains of more than 10 points on GSM-Hard, SVAMP, ASDIV, Bamboogle, and StrategyQA.The authors describe these results as evidence of the method’s general effectiveness.Source: §4.2 · pp. 5–6
• Self-consistency further enhanced contrastive CoT, particularly on AQuA, where the standalone gain was 4.0% and self-consistency added 14.2%.The passage describes the 14.2% as an additional improvement.Source: §4.2 · p. 6

Large Language Models; Chain of Thought

• The discussion describes language models’ prompt-based generalization while noting that scaling alone has not ensured strong performance on challenging arithmetic and factual reasoning tasks.It also reviews chain-of-thought work on automatic demonstrations, zero-shot prompting, problem decomposition, and executable programs.Source: §5 · p. 6

Related Work

• The paper positions contrastive chain-of-thought prompting as a response to limited understanding of its underlying mechanism and as a reason to rethink the process using valid and invalid demonstrations.The related-work discussion connects this proposal to earlier chain-of-thought research (Turpin et al., 2023; Feng et al., 2023).Source: §5 · p. 6

Learning from Negative Examples

• Prior work on contrastive learning and reinforcement learning from human feedback also uses positive and negative examples, motivating the paper’s approach to reasoning demonstrations.The passage cites contrastive learning and RLHF (Ouyang et al., 2022; Christiano et al., 2017).Source: §5 · p. 6

• Contrastive chain of thought is reported as a general enhancement of chain-of-thought prompting across several reasoning tasks.Source: §6 · p. 6
• A preliminary study found that providing valid and invalid reasoning demonstrations contrastively greatly improves language-model reasoning.Source: §6 · p. 6
• The method automatically constructs contrastive demonstrations from existing rationales to avoid manually annotating invalid rationales.Source: §6 · p. 6