Reasoning & CoT

Encourage step-by-step reasoning through intermediate thinking steps.

Add “Let’s think step by step” without examples for reasoning tasks.

Automatically generate diverse reasoning demonstrations with clustering.

Provide both correct and incorrect reasoning examples for contrast.

Triple-phase reasoning: analysis, structured thinking, summarization.

Symbolic reasoning combined with natural language for verifiable answers.

Select the most complex reasoning paths for harder problems.

Tabular format to organize step-by-step reasoning in columns.

Solve problems by generating and applying relevant analogies.

Abstract to higher-level concepts before solving specific problems.

Segment and reason through lengthy context systematically.

Identify uncertain examples and annotate them for better demonstrations.

Pre-think across unlabeled data before tackling target problems.

Work backward from conclusions to verify reasoning validity.

Apply chain-of-thought selectively based on model confidence for cost-optimal reasoning.

Embed algorithmic search patterns like DFS and BFS directly into prompts for single-call solution-space exploration.

Maintain reusable thought-templates from past problem-solving for template-guided reasoning.

Build recursive explanation trees and use logical consistency to verify commonsense reasoning.

Show intermediate computation steps in a dedicated workspace before producing final answers.

Alternate between selecting relevant premises and drawing conclusions for structured logical reasoning.

Solve analogous problems first and propagate insights to refine the target problem’s solution.

Replace natural language with compact symbolic representations for spatial and structural reasoning.

Integrate formal logic rules and symbolic expressions with chain-of-thought for verifiable reasoning.

Route between fast intuitive responses and slow deliberate reasoning based on task complexity.

Generate structured evidence triples with explanation hints for transparent, verifiable reasoning.

Reason with abstract placeholders first, then fill in concrete values using domain tools.

Combine reinforcement learning and Monte Carlo Tree Search for flexible, efficient reasoning.

Reason in terse shorthand drafts instead of verbose steps, cutting tokens while keeping accuracy.

A router picks one of three terse reasoning sketches per question to cut tokens while fitting the task.

Estimate a token budget per problem and prompt the model to reason within it, cutting wasted tokens.

A single model builds a reasoning DAG via proposer, critic, and summarizer roles with a validator.

Extract propositions, expand them with logic rules, and inject the derived facts back as context.

The model discovers a task-specific reasoning structure from atomic modules, then follows it.

Break complex problems into smaller, manageable parts with tree and graph approaches.

Critique, verify, and iteratively refine AI outputs for better accuracy.

Combine multiple reasoning paths for more reliable results.