"MTMT: Consolidating Multiple Thinking Modes to Form a Thought Tree for Strengthening LLM"

Tree-based thinking framework helps LLMs break down complex problems into manageable chunks.

MTMT introduces a novel framework that enhances LLM reasoning by combining multiple thinking modes into a tree structure, simulating both fast intuitive and slow deliberate cognitive processes to improve complex problem-solving capabilities.

-----

https://arxiv.org/abs/2412.03987

🤔 Original Problem:

LLMs struggle with complex logical reasoning and multi-step problem-solving tasks, showing limitations in handling intricate problems that require deep analytical thinking.

-----

🔧 Solution in this Paper:

→ MTMT constructs a thought tree that integrates various cognitive processes including association, counterfactual thinking, task decomposition, and comparison.

→ The system uses a state machine approach to select thinking modes based on current state and generates corresponding prompts.

→ Each thinking node in the tree processes existing information and determines the next strategy through perplexity threshold mechanisms.

→ The framework enables recursive interaction between fast (System 1) and slow (System 2) thinking approaches.

-----

💡 Key Insights:

→ Integration of multiple thinking modes significantly enhances LLM's reasoning capabilities

→ Thought tree structure improves answer interpretability by tracking reasoning paths

→ Perplexity threshold controls node generation and regeneration effectively

→ The approach is generalizable across different types of problems without modification

-----

📊 Results:

→ Achieved 44.0% accuracy on GPQA dataset without external knowledge

→ Demonstrated significant performance improvements across GPQA, TruthfulQA, and GSM8K datasets

→ Showed effective generalization across various datasets without prompt adjustments