"Chain-of-Thought in Large Language Models: Decoding, Projection, and Activation"

Inside the black box of Chain-of-Thought (CoT). How CoT helps LLMs combine format patterns with deep knowledge.

This paper investigates how Chain-of-Thought (CoT) prompting works internally in LLMs by examining three aspects: decoding behavior, projection space changes, and neuron activation patterns. The research reveals key mechanisms behind CoT's effectiveness in enhancing reasoning capabilities.

-----

https://arxiv.org/abs/2412.03944

🤔 Original Problem:

While CoT prompting improves LLM reasoning, we don't understand how it actually works internally. Previous research focused on external factors like step count and prompt design, but the internal mechanisms remained unclear.

-----

🔍 Solution in this Paper:

→ The researchers analyzed three distinct aspects of LLM behavior during CoT prompting.

→ They tracked how models imitate exemplar formats while integrating them with question understanding.

→ They examined changes in token probability distributions during generation.

→ They measured neuron activation patterns in the model's final layers.

→ They used multiple datasets across arithmetic, commonsense, and symbolic reasoning tasks.

-----

💡 Key Insights:

→ Models don't just copy CoT formats - they combine format imitation with deep understanding

→ CoT causes token probability fluctuations during generation but leads to more focused final outputs

→ CoT activates a broader range of neurons, suggesting more extensive knowledge retrieval

→ Better format imitation correlates with higher accuracy in transfer learning tests

-----

📊 Results:

→ CoT prompts achieve 0.85 average token probability vs 0.95 for standard prompts

→ CoT activates significantly more neurons in final model layers

→ Models maintain consistent performance even when CoT prompts transfer across different tasks