Focused Chain Of Thought

Summary

Focused Chain-of-Thought (F-CoT) is a training-free prompting strategy that improves inference efficiency by separating information extraction from reasoning. The method first organizes relevant information from a query into a structured, compact format, then guides the model to reason exclusively over this organized context rather than the original verbose question.

Core Technique

The approach operates in two stages:

Stage 1 - Information Organization: Parse the original query and extract relevant facts, constraints, and context. Format these into a clean, structured representation (tables, lists, key-value pairs). This extraction removes verbal noise and redundancy.

Stage 2 - Focused Reasoning: Present the organized information to the model alongside a query, asking it to reason exclusively from the structured context. The model no longer needs to parse or filter through verbose original text.

This two-stage separation enables the model to focus computation on reasoning rather than information comprehension.

Implementation

Extraction template: Design a format for your domain. Examples:

• For math: "Given: [list constraints], Find: [objective], Constraints: [list conditions]"
• For code review: "Files changed: [list], Key patterns: [list], Testing requirements: [list]"
• For analysis: "Data points: [structured table], Hypothesis: [statement], Constraints: [list]"

Structured representation: Convert original text into tables, bullet points, or key-value pairs. Each element should be a complete, self-contained fact.

Focused prompt: Chain-of-thought prompt operating only on structured context:

Context: [structured information]
Task: [reasoning goal]
Reasoning:

When to Use

• Tasks with naturally verbose inputs (documents, specifications, chat histories)
• Scenarios where token budget is constrained (long-context applications)
• Reasoning tasks benefiting from organized context representation
• Applications where input clarity directly impacts reasoning quality

When NOT to Use

• Tasks where implicit context hidden in prose is important to reasoning
• Scenarios where information extraction itself is complex or error-prone
• Applications with highly unstructured or ambiguous inputs
• Tasks benefiting from full original context for semantic understanding

Key References

• Chain-of-thought prompting and reasoning strategies
• Information extraction and structured formatting techniques
• Token efficiency and inference cost optimization