Implicit Chain of Thought in Large Language Models

This ongoing research project explores training language models that can reason internally using hidden states instead of articulating all reasoning steps explicitly, similar to human internal thought processes.

Project Overview

Objective: Develop language models capable of implicit reasoning without explicit step-by-step verbalization

Methodology:

• Training LMs to reason using internal hidden states
• Focus on simultaneous 2-digit multiplication tasks
• Knowledge distillation from explicit reasoning models
• Evaluation on mathematical reasoning benchmarks

Key Research Questions:

• Can language models learn to reason without explicitly verbalizing each step?
• How do implicit reasoning mechanisms compare to explicit Chain of Thought?
• What are the advantages of internal reasoning for complex tasks?

Current Work:

• Working on simultaneous 2-digit multiplication using Implicit Chain of Thought (ICoT)
• Exploring how models can develop internal reasoning mechanisms
• Analyzing the efficiency and accuracy trade-offs

Significance:

• More efficient reasoning without verbose explanations
• Better understanding of how LLMs process complex reasoning
• Potential for faster inference in production systems
• Insights into developing more human-like AI reasoning

Supervisor: Prof. Yuntian Deng
Institution: Deng’s Research Group, University of Waterloo
Duration: Fall 2024 - Present
Status: Ongoing Research

Recommended citation: Deng, Y., et al. (2024). "Implicit Chain of Thought Reasoning via Knowledge Distillation." arXiv preprint.
Download Paper