This Machine Learning Research Discusses Understanding the Reasoning Ability of Language Models from the Perspective of Reasoning Paths Aggregation

Large Language Models (LLMs) have proven to be exceptionally good at handling complicated reasoning problems in recent times. These tasks include solving mathematical puzzles, applying logic to solve difficulties, and solving challenges involving world knowledge without explicit fine-tuning. Researchers have been trying to answer the question of what role pre-training has in establishing reasoning capacities through next-token prediction.

In recent research, a team of researchers has focused on comprehending the emergence of reasoning ability, basically, the ability to infer new information from previously acquired knowledge. Intensive pre-training is how LLMs acquire emergent abilities. The research mainly aims to study the contribution of pre-training data to language model reasoning.

The team has shared that this study takes a Bayesian approach to explain why reasoning abilities in LLMs can be unlocked through next-token-prediction pre-training. According to the hypothesis, LLMs can use the next-token prediction target to collect indirect reasoning paths observed during pre-training. Reasoning routes can be considered textual arguments linking two concepts in real-world situations. According to the idea, LLMs might use these reasoning channels to hop from one notion to another during inference, which can result in the creation of chain-of-thought (CoT) solutions or silent reasoning that produces no explicit outputs.

Previous research has emphasized the significance of localized structures in the connections between variables in training data, particularly for CoT reasoning. However, this study asserts that when a reasoning path associates two concepts, they are likely to co-occur in the data, generating a graph-like localized structure. 

The study has focussed on two common forms of reasoning, mathematical and logical, in order to test these theories. The analysis is about reasoning over knowledge graphs (KGs) with random walk paths developed during pre-training for logical reasoning. The work has shown that compared to traditional path ranking algorithms, an LM pre-trained on random walk paths from a KG can accurately infer missing related links.

The study has addressed the problem of solving math word problems (MWPs) for mathematical reasoning. The method uses pre-existing CoT training data to create random walk reasoning paths rather than starting from scratch when pre-training an LM. The LM is then trained using next-token prediction along these pathways. The team has shared that the results from experiments on several MWP datasets have consistently outperformed standard-guided fine-tuning. 

The team has summarized their primary contributions as follows.

1. Validation of the Weighted Random Walk Hypothesis: The findings have indicated that the idea of combining weighted random walk reasoning paths provides a reasonable explanation for how language models learn to reason. 

2. This method works well for both mathematical reasoning using math word problems and logical reasoning using knowledge graphs, demonstrating its versatility as a means of comprehending LM reasoning.

3. Effective Use of Unlabeled Reasoning Paths: The findings have shown that LMs may use unlabeled reasoning paths well, highlighting the possibility of incorporating the random walk idea into LMs’ continuous pre-training procedures. This shows that using such a method can greatly improve the models’ capacity to carry out multi-step reasoning tasks in practical settings.

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 and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

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

Don’t Forget to join our Telegram Channel