Chain-of-Thought Prompting

The model is explicitly instructed to output a sequence of natural language reasoning steps—a 'cognitive trace'—before stating the final answer. This transforms a single inference call into a multi-step logical process.

• Decomposition: Breaks complex legal questions into manageable sub-problems
• Transparency: Makes the model's 'thinking' auditable, a key requirement for legal tech
• Error Propagation: A mistake in an early step can cascade, requiring careful prompt design

Chain-of-thought reasoning is not explicitly programmed but emerges as a capability in sufficiently large language models. It is generally ineffective on smaller models, where it can degrade performance.

• Scale Threshold: Typically emerges reliably in models with 100B+ parameters
• Zero-Shot CoT: Simply adding 'Let's think step by step' can trigger the behavior without examples
• Few-Shot CoT: Providing exemplars of decomposed reasoning yields more robust results on complex legal tasks

CoT prompting dramatically improves performance on tasks requiring precise, multi-hop manipulation of symbols, numbers, and logical constraints—common in statutory analysis.

• Date Calculation: Accurately computing deadlines by chaining temporal logic steps
• Damage Calculation: Performing multi-step arithmetic for statutory damages or interest
• Rule Chaining: Applying a sequence of conditional tests from a statute to a fact pattern

The primary risk is a 'faithful reasoning' failure, where the model generates a plausible-sounding but logically invalid chain of thought that leads to an incorrect conclusion.

• Factual Hallucination: Inventing a case citation within a reasoning step
• Logical Leap: Skipping a necessary inferential step, creating a non-sequitur
• Mitigation: Pairing CoT with Chain-of-Verification or Self-Consistency to validate the reasoning path

Provides the model with a small number of input-output examples within the prompt to guide behavior. In legal contexts, this means showing the model 2-3 examples of a completed IRAC analysis before asking it to analyze a new fact pattern. Unlike Chain-of-Thought, few-shot relies on pattern matching rather than explicit reasoning instruction, but the two techniques are often combined for maximum accuracy on complex statutory interpretation tasks.

A decoding strategy that generates multiple independent reasoning paths for a single query and selects the most frequent conclusion. For legal applications like case outcome prediction, this means running Chain-of-Thought prompting 5-10 times with a non-zero temperature and taking the majority verdict. This technique dramatically reduces variance on tasks with a definitive correct answer, such as identifying the controlling precedent in a jurisdiction split.

Enables exploration of multiple concurrent reasoning paths with strategic backtracking. For complex legal strategy, the model might explore:

• Path A: Argue under strict liability theory
• Path B: Argue under negligence theory
• Evaluate both paths
• Backtrack from dead-ends This is essential for tasks like multi-jurisdictional analysis where different legal frameworks must be evaluated in parallel before selecting the strongest argument.

A self-fact-checking technique where the model generates an initial response, then drafts independent verification questions and answers them before delivering the final output. In legal AI, this means the model might state a rule, then ask itself 'Does this case actually stand for that proposition?' and verify against its training data. This directly addresses the citation hallucination problem endemic to legal language models.

The capability to generate responses in a predefined machine-readable format like JSON. When combined with Chain-of-Thought, the reasoning can be placed in a rationale field while the conclusion occupies a separate holding field. This is critical for integrating legal reasoning into downstream software pipelines such as contract review platforms or e-discovery tools that require parseable outputs.