We explore the mechanism of in-context learning and propose a hypothesis using locate-and-project method. In shallow layers, the features of demonstrations are merged into their corresponding labels, and the features of the input text are aggregated into the last token. In deep layers, in-context heads make great contributions. In each in-context head, the value-output matrix extracts the labels' features. Query and key matrices compute the attention weights between the input text and each demonstration. The larger the attention weight is, the more label information is transferred into the last token for predicting the next word. Query and key matrices can be regarded as two towers for learning the similarity metric between the input text and each demonstration. Based on this hypothesis, we explain why imbalanced labels and demonstration order affect predictions. We conduct experiments on GPT2 large, Llama 7B, 13B and 30B. The results can support our analysis. Overall, our study provides a new method and a reasonable hypothesis for understanding the mechanism of in-context learning. Our code will be released on github.

We find the location of factual knowledge in large language models by exploring the residual stream and analyzing subvalues in vocabulary space. We find the reason why subvalues have human-interpretable concepts when projecting into vocabulary space. The before-softmax values of subvalues are added by an addition function, thus the probability of top tokens in vocabulary space will increase. Based on this, we find using log probability increase to compute the significance of layers and subvalues is better than probability increase, since the curve of log probability increase has a linear monotonically increasing shape. Moreover, we calculate the inner products to evaluate how much a feed-forward network (FFN) subvalue is activated by previous layers. Base on our methods, we find where factual knowledge is stored. Specifically, attention layers store "Paris is related to France". FFN layers store "Paris is a capital/city", activated by attention subvalues related to "capital". We leverage our method on Baevski-18, GPT2 medium, Llama-7B and Llama-13B. Overall, we provide a new method for understanding the mechanism of transformers. We will release our code on github.