Zeping Yu

Hengyuan Zhang, Zhihao Zhang, Mingyang Wang et al.

Mechanistic Interpretability (MI) has emerged as a vital approach to demystify the opaque decision-making of Large Language Models (LLMs). However, existing reviews primarily treat MI as an observational science, summarizing analytical insights while lacking a systematic framework for actionable intervention. To bridge this gap, we present a practical survey structured around the pipeline: "Locate, Steer, and Improve." We formally categorize Localizing (diagnosis) and Steering (intervention) methods based on specific Interpretable Objects to establish a rigorous intervention protocol. Furthermore, we demonstrate how this framework enables tangible improvements in Alignment, Capability, and Efficiency, effectively operationalizing MI as an actionable methodology for model optimization. The curated paper list of this work is available at https://github.com/rattlesnakey/Awesome-Actionable-MI-Survey.

Zeping Yu, Sophia Ananiadou

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.

Zhiwei Liu, Tianlin Zhang, Kailai Yang et al.

With the advent of social media, an increasing number of netizens are sharing and reading posts and news online. However, the huge volumes of misinformation (e.g., fake news and rumors) that flood the internet can adversely affect people's lives, and have resulted in the emergence of rumor and fake news detection as a hot research topic. The emotions and sentiments of netizens, as expressed in social media posts and news, constitute important factors that can help to distinguish fake news from genuine news and to understand the spread of rumors. This article comprehensively reviews emotion-based methods for misinformation detection. We begin by explaining the strong links between emotions and misinformation. We subsequently provide a detailed analysis of a range of misinformation detection methods that employ a variety of emotion, sentiment and stance-based features, and describe their strengths and weaknesses. Finally, we discuss a number of ongoing challenges in emotion-based misinformation detection based on large language models and suggest future research directions, including data collection (multi-platform, multilingual), annotation, benchmark, multimodality, and interpretability.

Zeping Yu, Sophia Ananiadou

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.

Zeping Yu, Sophia Ananiadou

Understanding the mechanisms behind Large Language Models (LLMs) is crucial for designing improved models and strategies. While recent studies have yielded valuable insights into the mechanisms of textual LLMs, the mechanisms of Multi-modal Large Language Models (MLLMs) remain underexplored. In this paper, we apply mechanistic interpretability methods to analyze the visual question answering (VQA) mechanisms in the first MLLM, Llava. We compare the mechanisms between VQA and textual QA (TQA) in color answering tasks and find that: a) VQA exhibits a mechanism similar to the in-context learning mechanism observed in TQA; b) the visual features exhibit significant interpretability when projecting the visual embeddings into the embedding space; and c) Llava enhances the existing capabilities of the corresponding textual LLM Vicuna during visual instruction tuning. Based on these findings, we develop an interpretability tool to help users and researchers identify important visual locations for final predictions, aiding in the understanding of visual hallucination. Our method demonstrates faster and more effective results compared to existing interpretability approaches. Code: \url{https://github.com/zepingyu0512/llava-mechanism}

Zeping Yu, Sophia Ananiadou

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.

Zeping Yu, Yonatan Belinkov, Sophia Ananiadou

We investigate how large language models 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 four LLMs, back attention improves accuracy on five reasoning datasets, demonstrating its effectiveness in enhancing latent multi-hop reasoning ability.

Zeping Yu, Sophia Ananiadou

Large language models (LLMs) often exhibit gender bias, posing challenges for their safe deployment. Existing methods to mitigate bias lack a comprehensive understanding of its mechanisms or compromise the model's core capabilities. To address these issues, we propose the CommonWords dataset, to systematically evaluate gender bias in LLMs. Our analysis reveals pervasive bias across models and identifies specific neuron circuits, including gender neurons and general neurons, responsible for this behavior. Notably, editing even a small number of general neurons can disrupt the model's overall capabilities due to hierarchical neuron interactions. Based on these insights, we propose an interpretable neuron editing method that combines logit-based and causal-based strategies to selectively target biased neurons. Experiments on five LLMs demonstrate that our method effectively reduces gender bias while preserving the model's original capabilities, outperforming existing fine-tuning and editing approaches. Our findings contribute a novel dataset, a detailed analysis of bias mechanisms, and a practical solution for mitigating gender bias in LLMs.

Zeping Yu, Sophia Ananiadou

Although multimodal large language models (MLLMs) have achieved impressive performance, the multimodal instruction tuning stage often causes catastrophic forgetting of the base LLM's language ability, even in strong models like Llama3. To address this, we propose Locate-then-Merge, a training-free parameter fusion framework that first locates important parameters and then selectively merges them. We further introduce Neuron-Fusion, a neuron-level strategy that preserves the influence of neurons with large parameter shifts--neurons likely responsible for newly acquired visual capabilities--while attenuating the influence of neurons with smaller changes that likely encode general-purpose language skills. This design enables better retention of visual adaptation while mitigating language degradation. Experiments on 13 benchmarks across both language and visual tasks show that Neuron-Fusion consistently outperforms existing model merging methods. Further analysis reveals that our method effectively reduces context hallucination in generation.

Zeping Yu, Gongshen Liu

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 largescale sentiment analysis datasets show that SRNNs achieve better performance than standard RNNs.