We investigate how large language models (LLMs) perform latent multi-hop reasoning in prompts like "Wolfgang Amadeus Mozart's mother's spouse is".To analyze this process, we introduce logit flow, an interpretability method that traces how logits propagate across layers and positions toward the final prediction.Using logit flow, we identify four distinct stages in single-hop knowledge prediction: (A) entity subject enrichment, (B) entity attribute extraction, (C) relation subject enrichment, and (D) relation attribute extraction.Extending this analysis to multi-hop reasoning, we find that failures often stem from the relation attribute extraction stage, where conflicting logits reduce prediction accuracy.To address this, we propose back attention, a novel mechanism that enables lower layers to leverage higher-layer hidden states from different positions during attention computation.With back attention, a 1-layer transformer achieves the performance of a 2-layer transformer.Applied to five LLMs, back attention improves accuracy on five reasoning datasets, demonstrating its effectiveness in enhancing latent multi-hop reasoning ability.

Identifying important neurons for final predictions is essential for understanding the mechanisms of large language models.Due to computational constraints, current attribution techniques struggle to operate at neuron level.In this paper, we propose a static method for pinpointing significant neurons.Compared to seven other methods, our approach demonstrates superior performance across three metrics.Additionally, since most static methods typically only identify "value neurons" directly contributing to the final prediction, we propose a method for identifying "query neurons" which activate these "value neurons".Finally, we apply our methods to analyze six types of knowledge across both attention and feed-forward network (FFN) layers.Our method and analysis are helpful for understanding the mechanisms of knowledge storage and set the stage for future research in knowledge editing.The code is available on https://github.com/ zepingyu0512/neuron-attribution.

We investigate the mechanism of in-context learning (ICL) on sentence classification tasks with semantically-unrelated labels ("foo"/"bar").We find intervening in only 1% heads (named "in-context heads") significantly affects ICL accuracy from 87.6% to 24.4%.To understand this phenomenon, we analyze the value-output vectors in these heads and discover that the vectors at each label position contain substantial information about the corresponding labels.Furthermore, we observe that the prediction shift from "foo" to "bar" is due to the respective reduction and increase in these heads' attention scores at "foo" and "bar" positions.Therefore, we propose a hypothesis for ICL: in in-context heads, the value-output matrices extract label features, while the query-key matrices compute the similarity between the features at the last position and those at each label position.The query and key matrices can be considered as two towers that learn the similarity metric between the last position's features and each demonstration at label positions.Using this hypothesis, we explain the majority label bias and recency bias in ICL and propose two methods to reduce these biases by 22% and 17%, respectively. Shallow Layers France

We find arithmetic ability resides within a limited number of attention heads, with each head specializing in distinct operations.To delve into the reason, we introduce the Comparative Neuron Analysis (CNA) method, which identifies an internal logic chain consisting of four distinct stages from input to prediction: feature enhancing with shallow FFN neurons, feature transferring by shallow attention layers, feature predicting by arithmetic heads, and prediction enhancing among deep FFN neurons.Moreover, we identify the human-interpretable FFN neurons within both feature-enhancing and feature-predicting stages.These findings lead us to investigate the mechanism of LoRA, revealing that it enhances prediction probabilities by amplifying the coefficient scores of FFN neurons related to predictions.Finally, we apply our method in model pruning for arithmetic tasks and model editing for reducing gender bias.Code is on https://github.com/ zepingyu0512/arithmetic-mechanism.

Binary code similarity detection, whose goal is to detect similar binary functions without having access to the source code, is an essential task in computer security. Traditional methods usually use graph matching algorithms, which are slow and inaccurate. Recently, neural network-based approaches have made great achievements. A binary function is first represented as an control-flow graph (CFG) with manually selected block features, and then graph neural network (GNN) is adopted to compute the graph embedding. While these methods are effective and efficient, they could not capture enough semantic information of the binary code. In this paper we propose semantic-aware neural networks to extract the semantic information of the binary code. Specially, we use BERT to pre-train the binary code on one token-level task, one block-level task, and two graph-level tasks. Moreover, we find that the order of the CFG's nodes is important for graph similarity detection, so we adopt convolutional neural network (CNN) on adjacency matrices to extract the order information. We conduct experiments on two tasks with four datasets. The results demonstrate that our method outperforms the state-of-art models.

Binary source code matching, especially on function-level, has a critical role in the field of computer security. Given binary code only, finding the corresponding source code improves the accuracy and efficiency in reverse engineering. Given source code only, related binary code retrieval contributes to known vulnerabilities confirmation. However, due to the vast difference between source and binary code, few studies have investigated binary source code matching. Previously published studies focus on code literals extraction such as strings and integers, then utilize traditional matching algorithms such as the Hungarian algorithm for code matching. Nevertheless, these methods have limitations on function-level, because they ignore the potential semantic features of code and a lot of code lacks sufficient code literals. Also, these methods indicate a need for expert experience for useful feature identification and feature engineering, which is timeconsuming. This paper proposes an end-to-end cross-modal retrieval network for binary source code matching, which achieves higher accuracy and requires less expert experience. We adopt Deep Pyramid Convolutional Neural Network (DPCNN) for source code feature extraction and Graph Neural Network (GNN) for binary code feature extraction. We also exploit neural network-based models to capture code literals, including strings and integers. Furthermore, we implement "norm weighted sampling" for negative sampling. We evaluate our model on two datasets, where it outperforms other methods significantly.

User modeling is an essential task for online recommender systems. In the past few decades, collaborative filtering (CF) techniques have been well studied to model users' long term preferences. Recently, recurrent neural networks (RNN) have shown a great advantage in modeling users' short term preference. A natural way to improve the recommender is to combine both long-term and short-term modeling. Previous approaches neglect the importance of dynamically integrating these two user modeling paradigms. Moreover, users' behaviors are much more complex than sentences in language modeling or images in visual computing, thus the classical structures of RNN such as Long Short-Term Memory (LSTM) need to be upgraded for better user modeling. In this paper, we improve the traditional RNN structure by proposing a time-aware controller and a content-aware controller, so that contextual information can be well considered to control the state transition. We further propose an attention-based framework to combine users' long-term and short-term preferences, thus users' representation can be generated adaptively according to the specific context. We conduct extensive experiments on both public and industrial datasets. The results demonstrate that our proposed method outperforms several state-of-art methods consistently.

Recurrent neural networks have achieved great success in many NLP tasks. However, they have difficulty in parallelization because of the recurrent structure, so it takes much time to train RNNs. In this paper, we introduce sliced recurrent neural networks (SRNNs), which could be parallelized by slicing the sequences into many subsequences. SRNNs have the ability to obtain high-level information through multiple layers with few extra parameters. We prove that the standard RNN is a special case of the SRNN when we use linear activation functions. Without changing the recurrent units, SRNNs are 136 times as fast as standard RNNs and could be even faster when we train longer sequences. Experiments on six large-scale sentiment analysis datasets show that SRNNs achieve better performance than standard RNNs.

In a multi-terminal high-voltage direct current based on voltage source converter, the proactive hybrid HVDC breaker is an important component in DC grids to isolate a DC fault. From the DC breaker design point, it is crucial to realize a fast fault current breaking. Hybrid HVDC breaker based on delay slope method for breaking signal has more fast action. Especially for overhead lines of DC grid, automatic reclosing control method is adopted in this paper. A wind farm ring topology model, 8-terminal HDC based on voltage source converter, is build on PSCAD/EMTDC simulation platform. Simulations are given to demonstrate effectiveness of the delay slope method and automatic reclosing control strategies. In last, it is pointed out the existing problems in automatic reclosing of Hybrid HVDC breaker.