Large language models have consistently struggled with complex reasoning tasks, such as mathematical problem-solving. Investigating the internal reasoning mechanisms of these models can help us design better model architectures and training strategies, ultimately enhancing their reasoning capability. In this study, we constructed a symbolic multi-step reasoning task to investigate the information propagation mechanisms in Transformer models when solving the task through direct answering and Chain-of-Thought (CoT) reasoning. We introduced the concept of buffer mechanism: the model stores various information in distinct buffers and selectively extracts it through the query-key matrix. We proposed a random matrix-based algorithm to enhance the model’s reasoning ability. This algorithm introduces only 132 trainable parameters, yet leads to significant performance improvements on 7 multi-step reasoning datasets, including PrOntoQA, LogicAsker, and LogicInference. These findings provide new insights into understanding the large language models.

This paper presents our endeavor for solving task11, NLPContributionGraph, of SemEval-2021. The purpose of the task was to extract triples from a paper in the Nature Language Processing field for constructing an Open Research Knowledge Graph. The task includes three sub-tasks: detecting the contribution sentences in papers, identifying scientific terms and predicate phrases from the contribution sentences; and inferring triples in the form of (subject, predicate, object) as statements for Knowledge Graph building. In this paper, we apply an ensemble of various fine-tuned pre-trained language models (PLM) for tasks one and two. In addition, self-training methods are adopted for tackling the shortage of annotated data. For the third task, rather than using classic neural open information extraction (OIE) architectures, we generate potential triples via manually designed rules and develop a binary classifier to differentiate positive ones from others. The quantitative results show that we obtain the 4th, 2nd, and 2nd rank in three evaluation phases.