Microsoft AI Proposes CoT-Influx: A Novel Machine Learning Approach that Pushes the Boundary of Few-Shot Chain-of-Thoughts (CoT) Learning to Improve LLM Mathematical Reasoning

With a nuanced scope of application, because of the amount of information it has been exposed and trained to, Large Language Models (LLMs) have emerged as game changers in Artificial Intelligence (AI). However, there are still unexplored or less explored territories that sometimes need improvement. One such territory is its ability to reason mathematically. These models, particularly smaller ones like LLaMA, face challenges in math reasoning, which is a critical component of AI’s cognitive capabilities. The research community is tirelessly working towards optimizing Chain-of-Thought (CoT) prompts and fine-tuning LLMs to enhance their reasoning skills. Yet, the full potential of few-shot learning still needs to be explored.

Recent research has improved the reasoning capabilities of LLMs by enhancing CoT prompts and innovating CoT-based training data. Prompt compression methods have been explored to address the challenge of limited few-shot examples, but they must solve the problem effectively. Prompt retrieval methods optimize task performance by selecting high-quality few-shot examples, but they are sub-optimal for math reasoning and do not account for token redundancy. The accuracy of LLaMA2-7B reasoning decreases as the number of CoT examples exceeds token limits. LLMs with different capabilities favor CoT examples of varying complexities, but current retrieval methods do not consider this.

A research team from Hong Kong University and Microsoft has proposed CoT-Influx. This novel approach introduces a more effective use of few-shot learning to boost LLM math reasoning capabilities. Leveraging a coarse-to-fine pruning mechanism, CoT-Influx aims to maximize the input of effective and concise CoT examples within the confines of existing context windows. This approach allows for more helpful CoT examples and ensures that each example comprises informative tokens.

The development of CoT-Influx involved the creation of a specialized math reasoning dataset, MRD3, featuring problems that span over a wide range of difficulty levels and reasoning steps. This dataset is the foundation for training a specialized pruner tailored for math reasoning tasks. The pruner operates in two pivotal stages—initially selecting the quintessential CoT examples from a vast pool and subsequently pruning the superfluous tokens to conform to the original context window’s constraints. By adopting this dual-phase pruning strategy, CoT-Influx effectively doubles the context window’s capacity for useful CoT examples without incurring additional computational overhead or complexity.

The effectiveness of CoT-Influx is proven through rigorous testing, showing a significant boost in LLMs’ math-solving abilities. Applied to various LLaMA models over five math datasets, CoT-Influx led to considerable accuracy improvements. A key highlight is the LLaMA2-70B model with CoT-Influx surpassing the GPT-3.5 and larger models on the GSM8K dataset by a remarkable 2.5%. Moreover, across other datasets like AddSub and Multiarith, CoT-Influx enabled models to achieve top performance, underscoring its critical role in advancing LLMs’ mathematical reasoning capabilities.

In conclusion, the study introduces CoT-Influx, a method that significantly enhances the math reasoning capabilities of LLMs like LLaMA. By efficiently pruning and utilizing math-related examples, CoT-Influx allows these models to achieve higher accuracy on challenging datasets, such as GSM8K, AddSub, and Multiarith. This advancement marks a significant step forward and opens up new possibilities for applying LLMs to solve complex mathematical problems, indicating a promising direction for future research in AI reasoning and learning efficiency.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. Join our Telegram Channel, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our 39k+ ML SubReddit