机制可解释性

Retrieval-Augmented Generation (RAG) models are designed to incorporate external knowledge, reducing hallucinations caused by insufficient parametric (internal) knowledge. However, even with accurate and relevant retrieved content, RAG models can still produce hallucinations by generating outputs that conflict with the retrieved information. Detecting such hallucinations requires disentangling how Large Language Models (LLMs) balance external and parametric knowledge. Current detection methods often focus on one of these mechanisms or without decoupling their intertwined effects, making accurate detection difficult. In this paper, we investigate the internal mechanisms behind hallucinations in RAG scenarios. We discover hallucinations occur when the **Knowledge FFNs** in LLMs overemphasize parametric knowledge in the residual stream, while **Copying Heads** fail to effectively retain or integrate external knowledge from retrieved content. Based on these findings, we propose **ReDeEP**, a novel method that detects hallucinations by decoupling LLM’s utilization of external context and parametric knowledge. Our experiments show that ReDeEP significantly improves RAG hallucination detection accuracy. Additionally, we introduce AARF, which mitigates hallucinations by modulating the contributions of Knowledge FFNs and Copying Heads.

In-context Learning (ICL) is an emerging few-shot learning paradigm on Language Models (LMs) with inner mechanisms un-explored. There are already existing works describing the inner processing of ICL, while they struggle to capture all the inference phenomena in large language models. Therefore, this paper proposes a comprehensive circuit to model the inference dynamics and try to explain the observed phenomena of ICL. In detail, we divide ICL inference into 3 major operations: (1) Input Text Encode: LMs encode every input text (in the demonstrations and queries) into linear representation in the hidden states with sufficient information to solve ICL tasks. (2) Semantics Merge: LMs merge the encoded representations of demonstrations with their corresponding label tokens to produce joint representations of labels and demonstrations. (3) Feature Retrieval and Copy: LMs search the joint representations of demonstrations similar to the query representation on a task subspace, and copy the searched representations into the query. Then, language model heads capture these copied label representations to a certain extent and decode them into predicted labels. Through careful measurements, the proposed inference circuit successfully captures and unifies many fragmented phenomena observed during the ICL process, making it a comprehensive and practical explanation of the ICL inference process. Moreover, ablation analysis by disabling the proposed steps seriously damages the ICL performance, suggesting the proposed inference circuit is a dominating mechanism. Additionally, we confirm and list some bypass mechanisms that solve ICL tasks in parallel with the proposed circuit.

The ability to follow instructions is crucial for numerous real-world applications of language models. In pursuit of deeper insights and more powerful capabilities, we derive instruction-specific vector representations from language models and use them to steer models accordingly. These vectors are computed as the difference in activations between inputs with and without instructions, enabling a modular approach to activation steering. We demonstrate how this method can enhance model adherence to constraints such as output format, length, and word inclusion, providing inference-time control over instruction following. Our experiments across four models demonstrate how we can use the activation vectors to guide models to follow constraints even without explicit instructions and to enhance performance when instructions are present. Additionally, we explore the compositionality of activation steering, successfully applying multiple instructions simultaneously. Finally, we demonstrate that steering vectors computed on instruction-tuned models can transfer to improve base models. Our findings demonstrate that activation steering offers a practical and scalable approach for fine-grained control in language generation. Our code and data are available at https://github.com/microsoft/llm-steer-instruct.

Recent advances in efficient sequence modeling have led to attention-free layers, such as Mamba, RWKV, and various gated RNNs, all featuring sub-quadratic complexity in sequence length and excellent scaling properties, enabling the construction of a new type of foundation models. In this paper, we present a unified view of these models, formulating such layers as implicit causal self-attention layers. The formulation includes most of their sub-components and is not limited to a specific part of the architecture. The framework compares the underlying mechanisms on similar grounds for different layers and provides a direct means for applying explainability methods. Our experiments show that our attention matrices and attribution method outperform an alternative and a more limited formulation that was recently proposed for Mamba. For the other architectures for which our method is the first to provide such a view, our method is effective and competitive in the relevant metrics compared to the results obtained by state-of-the-art Transformer explainability methods. Our code is publicly available.

Large language models (LLMs) demonstrate remarkable emergent abilities to perform in-context learning across various tasks, including time series forecasting. This work investigates LLMs' ability to estimate probability density functions (PDFs) from data observed in-context; such density estimation (DE) is a fundamental task underlying many probabilistic modeling problems. We leverage the Intensive Principal Component Analysis (InPCA) to visualize and analyze the in-context learning dynamics of LLaMA-2 models. Our main finding is that these LLMs all follow similar learning trajectories in a low-dimensional InPCA space, which are distinct from those of traditional density estimation methods like histograms and Gaussian kernel density estimation (KDE). We interpret the LLaMA in-context DE process as a KDE with an adaptive kernel width and shape. This custom kernel model captures a significant portion of LLaMA's behavior despite having only two parameters. We further speculate on why LLaMA's kernel width and shape differs from classical algorithms, providing insights into the mechanism of in-context probabilistic reasoning in LLMs. Our codebase, along with a 3D visualization of an LLM's in-context learning trajectory, is publicly available at https://github.com/AntonioLiu97/LLMICL_inPCA.

Large language models (LLMs) are vulnerable to adversarial attacks that can elicit harmful responses. Defending against such attacks remains challenging due to the opacity of jailbreaking mechanisms and the high computational cost of training LLMs robustly. We demonstrate that adversarial attacks share a universal mechanism for circumventing LLM safeguards that works by ablating a dimension in the residual stream embedding space called the refusal feature. We further show that the operation of refusal feature ablation (RFA) approximates the worst-case perturbation of offsetting model safety. Based on these findings, we propose Refusal Feature Adversarial Training (ReFAT), a novel algorithm that efficiently performs LLM adversarial training by simulating the effect of input-level attacks via RFA. Experiment results show that ReFAT significantly improves the robustness of three popular LLMs against a wide range of adversarial attacks, with considerably less computational overhead compared to existing adversarial training methods.

Large language models (LLMs) have demonstrated the ability to generate text that realistically reflects a range of different subjective human perspectives. This paper studies how LLMs are seemingly able to reflect more liberal versus more conservative viewpoints among other political perspectives in American politics. We show that LLMs possess linear representations of political perspectives within activation space, wherein more similar perspectives are represented closer together. To do so, we probe the attention heads across the layers of three open transformer-based LLMs (Llama-2-7b-chat, Mistral-7b-instruct, Vicuna-7b). We first prompt models to generate text from the perspectives of different U.S. lawmakers. We then identify sets of attention heads whose activations linearly predict those lawmakers' DW-NOMINATE scores, a widely-used and validated measure of political ideology. We find that highly predictive heads are primarily located in the middle layers, often speculated to encode high-level concepts and tasks. Using probes only trained to predict lawmakers' ideology, we then show that the same probes can predict measures of news outlets' slant from the activations of models prompted to simulate text from those news outlets. These linear probes allow us to visualize, interpret, and monitor ideological stances implicitly adopted by an LLM as it generates open-ended responses. Finally, we demonstrate that by applying linear interventions to these attention heads, we can steer the model outputs toward a more liberal or conservative stance. Overall, our research suggests that LLMs possess a high-level linear representation of American political ideology and that by leveraging recent advances in mechanistic interpretability, we can identify, monitor, and steer the subjective perspective underlying generated text.

Outliers have been widely observed in Large Language Models (LLMs), significantly impacting model performance and posing challenges for model compression. Understanding the functionality and formation mechanisms of these outliers is critically important. Existing works, however, largely focus on reducing the impact of outliers from an algorithmic perspective, lacking an in-depth investigation into their causes and roles. In this work, we provide a detailed analysis of the formation process, underlying causes, and functions of outliers in LLMs. We define and categorize three types of outliers—activation outliers, weight outliers, and attention outliers—and analyze their distributions across different dimensions, uncovering inherent connections between their occurrences and their ultimate influence on the attention mechanism. Based on these observations, we hypothesize and explore the mechanisms by which these outliers arise and function, demonstrating through theoretical derivations and experiments that they emerge due to the self-attention mechanism's softmax operation. These outliers act as implicit context-aware scaling factors within the attention mechanism. As these outliers stem from systematic influences, we term them systematic outliers. Our study not only enhances the understanding of Transformer-based LLMs but also shows that structurally eliminating outliers can accelerate convergence and improve model compression. The code is avilable at \url{https://github.com/an-yongqi/systematic-outliers}.

Recent alignment algorithms such as direct preference optimization (DPO) have been developed to improve the safety of large language models (LLMs) by training these models to match human behaviors exemplified by preference data. However, these methods are both computationally intensive and lacking in controllability and transparency, inhibiting their widespread use. Furthermore, these tuning-based methods require large-scale preference data for training and are susceptible to noisy preference data. In this paper, we introduce a tuning-free alignment alternative, ProFS (Projection Filter for Subspaces), and demonstrate its effectiveness under the use case of toxicity reduction. Grounded on theory from factor analysis, ProFS is a sample-efficient model editing approach that identifies a toxic subspace in the model parameter space and reduces model toxicity by projecting away the detected subspace. The toxic subspace is identified by extracting preference data embeddings from the language model, and removing non-toxic information from these embeddings. We show that ProFS is more sample-efficient than DPO, further showcasing greater robustness to noisy data. Finally, we attempt to connect tuning based alignment with editing, by establishing both theoretical and empirical connections between ProFS and DPO, showing that ProFS can be interpreted as a denoised version of a single DPO step.

A mechanistic understanding of how MLPs do computation in deep neural net- works remains elusive. Current interpretability work can extract features from hidden activations over an input dataset but generally cannot explain how MLP weights construct features. One challenge is that element-wise nonlinearities introduce higher-order interactions and make it difficult to trace computations through the MLP layer. In this paper, we analyze bilinear MLPs, a type of Gated Linear Unit (GLU) without any element-wise nonlinearity that neverthe- less achieves competitive performance. Bilinear MLPs can be fully expressed in terms of linear operations using a third-order tensor, allowing flexible analysis of the weights. Analyzing the spectra of bilinear MLP weights using eigendecom- position reveals interpretable low-rank structure across toy tasks, image classifi- cation, and language modeling. We use this understanding to craft adversarial examples, uncover overfitting, and identify small language model circuits directly from the weights alone. Our results demonstrate that bilinear layers serve as an interpretable drop-in replacement for current activation functions and that weight- based interpretability is viable for understanding deep-learning models.

Position bias has proven to be a prevalent issue of modern language models (LMs), where the models prioritize content based on its position within the given context. This bias often leads to unexpected model failures and hurts performance, robustness, and reliability across various applications. A simple mechanistic analysis attributes the position bias to two components employed in nearly all state-of-the-art LMs: causal attention and position embedding. Based on the analyses, we propose to **eliminate** position bias (e.g., different retrieved documents' orders in QA affect performance) with a **training-free zero-shot** approach. Our method changes the causal attention to bidirectional attention between documents and utilizes model attention values to decide the relative orders of documents instead of using the order provided in input prompts, therefore enabling Position-INvariant inferencE (PINE) at the document level. By eliminating position bias, models achieve better performance and reliability in downstream tasks, including LM-as-a-judge, retrieval-augmented QA, molecule generation, and math reasoning. Notably, PINE is especially useful when adapting LMs for evaluating reasoning pairs: it consistently provides $8$ to $10$ percentage points performance gains, making Llama-3-70B-Instruct perform even better than GPT-4-0125-preview and GPT-4o-2024-08-06 on the RewardBench reasoning set.

Recent work has proposed that language models perform computation by manipulating one-dimensional representations of concepts ("features") in activation space. In contrast, we explore whether some language model representations may be inherently multi-dimensional. We begin by developing a rigorous definition of irreducible multi-dimensional features based on whether they can be decomposed into either independent or non-co-occurring lower-dimensional features. Motivated by these definitions, we design a scalable method that uses sparse autoencoders to automatically find multi-dimensional features in GPT-2 and Mistral 7B. These auto-discovered features include strikingly interpretable examples, e.g. $\textit{circular}$ features representing days of the week and months of the year. We identify tasks where these exact circles are used to solve computational problems involving modular arithmetic in days of the week and months of the year. Next, we provide evidence that these circular features are indeed the fundamental unit of computation in these tasks with intervention experiments on Mistral 7B and Llama 3 8B, and we examine the continuity of the days of the week feature in Mistral 7B. Overall, our work argues that understanding multi-dimensional features is necessary to mechanistically decompose some model behaviors.

Vision-Language Models (VLMs) are powerful tools for processing and understanding text and images. We study the processing of visual tokens in the language model component of LLaVA, a prominent VLM. Our approach focuses on analyzing the localization of object information, the evolution of visual token representations across layers, and the mechanism of integrating visual information for predictions. Through ablation studies, we demonstrated that object identification accuracy drops by over 70\% when object-specific tokens are removed. We observed that visual token representations become increasingly interpretable in the vocabulary space across layers, suggesting an alignment with textual tokens corresponding to image content. Finally, we found that the model extracts object information from these refined representations at the last token position for prediction, mirroring the process in text-only language models for factual association tasks. These findings provide crucial insights into how VLMs process and integrate visual information, bridging the gap between our understanding of language and vision models, and paving the way for more interpretable and controllable multimodal systems.

Sparse autoencoders (SAEs) are a promising approach to interpreting the internal representations of transformer language models. However, SAEs are usually trained separately on each transformer layer, making it difficult to use them to study how information flows across layers. To solve this problem, we introduce the multi-layer SAE (MLSAE): a single SAE trained on the residual stream activation vectors from every transformer layer. Given that the residual stream is understood to preserve information across layers, we expected MLSAE latents to 'switch on' at a token position and remain active at later layers. Interestingly, we find that individual latents are often active at a single layer for a given token or prompt, but the layer at which an individual latent is active may differ for different tokens or prompts. We quantify these phenomena by defining a distribution over layers and considering its variance. We find that the variance of the distributions of latent activations over layers is about two orders of magnitude greater when aggregating over tokens compared with a single token. For larger underlying models, the degree to which latents are active at multiple layers increases, which is consistent with the fact that the residual stream activation vectors at adjacent layers become more similar. Finally, we relax the assumption that the residual stream basis is the same at every layer by applying pre-trained tuned-lens transformations, but our findings remain qualitatively similar. Our results represent a new approach to understanding how representations change as they flow through transformers. We release our code to train and analyze MLSAEs at https://github.com/tim-lawson/mlsae.

Hallucinations in large language models are a widespread problem, yet the mechanisms behind whether models will hallucinate are poorly understood, limiting our ability to solve this problem. Using sparse autoencoders as an interpretability tool, we discover that a key part of these mechanisms is entity recognition, where the model detects if an entity is one it can recall facts about. Sparse autoencoders uncover meaningful directions in the representation space, these detect whether the model recognizes an entity, e.g. detecting it doesn't know about an athlete or a movie. This shows that models can have self-knowledge: internal representations about their own capabilities. These directions are causally relevant: capable of steering the model to refuse to answer questions about known entities, or to hallucinate attributes of unknown entities when it would otherwise refuse. We demonstrate that despite the sparse autoencoders being trained on the base model, these directions have a causal effect on the chat model's refusal behavior, suggesting that chat finetuning has repurposed this existing mechanism. Furthermore, we provide an initial exploration into the mechanistic role of these directions in the model, finding that they disrupt the attention of downstream heads that typically move entity attributes to the final token.

Balancing high performance with interpretability in increasingly powerful Transformer-based models remains a challenge. While mechanistic interpretability aims to specify neural network computations in explicit, pseudocode-like formats, existing methods often involve laborious manual analysis or struggle to fully elucidate learned internal algorithms. Recent efforts to build intrinsically interpretable models have introduced considerable expressivity and optimization challenges. This work introduces Adaptive Transformer Programs, an enhanced framework building upon RASP language and Transformer Programs to create more robust and interpretable models. The proposed method increases expressivity by redesigning two primary attention modules to improve categorical and numerical reasoning capabilities. To overcome optimization hurdles, we introduce a novel reparameterization scheme that enhances the exploration-exploitation trade-off during training. We validate our approach through extensive experiments on diverse tasks, including in-context learning, algorithmic problems (e.g., sorting and Dyck languages), and NLP benchmarks such as named entity recognition and text classification. Results demonstrate that Adaptive Transformer Programs substantially narrow the performance gap between black-box Transformers and interpretable models, enhancing transparency. This work advances the development of high-performing, transparent AI systems for critical applications, addressing crucial ethical concerns in AI development.

Transformers can under some circumstances generalize to novel problem instances whose constituent parts might have been encountered during training, but whose compositions have not. What mechanisms underlie this ability for compositional generalization? By reformulating multi-head attention as a hypernetwork, we reveal that a composable, low-dimensional latent code specifies key-query specific operations. We find empirically that this latent code is predictive of the subtasks the network performs on unseen task compositions, revealing that latent codes acquired during training are reused to solve unseen problem instances. To further examine the hypothesis that the intrinsic hypernetwork of multi-head attention supports compositional generalization, we ablate whether making the hypernetwork-generated linear value network nonlinear strengthens compositionality. We find that this modification improves compositional generalization on abstract reasoning tasks. In particular, we introduce a symbolic version of the Raven's Progressive Matrices human intelligence test, which gives us precise control over the problem compositions encountered during training and evaluation. We demonstrate on this task how scaling model size and data enables compositional generalization in transformers and gives rise to a functionally structured latent space.

We introduce refined variants of the Local Learning Coefficient (LLC), a measure of model complexity grounded in singular learning theory, to study the development of internal structure in transformer language models during training. By applying these refined LLCs (rLLCs) to individual components of a two-layer attention-only transformer, we gain novel insights into the progressive differentiation and specialization of attention heads. Our methodology reveals how attention heads differentiate into distinct functional roles over the course of training, analyzes the types of data these heads specialize to process, and discovers a previously unidentified multigram circuit. These findings demonstrate that rLLCs provide a principled, quantitative toolkit for developmental interpretability, which aims to understand models through their evolution across the learning process. This work advances the field of developmental interpretability by providing a mathematically rigorous approach to understanding neural networks through the lens of their learning process. More broadly, this work takes a step towards establishing the correspondence between data distributional structure, geometric properties of the loss landscape, learning dynamics, and emergent computational structures in neural networks.

Many proposed applications of neural networks in machine learning, cognitive/brain science, and society hinge on the feasibility of inner interpretability via circuit discovery. This calls for empirical and theoretical explorations of viable algorithmic options. Despite advances in the design and testing of heuristics, there are concerns about their scalability and faithfulness at a time when we lack understanding of the complexity properties of the problems they are deployed to solve. To address this, we study circuit discovery with classical and parameterized computational complexity theory: (1) we describe a conceptual scaffolding to reason about circuit finding queries in terms of affordances for description, explanation, prediction and control; (2) we formalize a comprehensive set of queries for mechanistic explanation, and propose a formal framework for their analysis; (3) we use it to settle the complexity of many query variants and relaxations of practical interest on multi-layer perceptrons. Our findings reveal a challenging complexity landscape. Many queries are intractable, remain fixed-parameter intractable relative to model/circuit features, and inapproximable under additive, multiplicative, and probabilistic approximation schemes. To navigate this landscape, we prove there exist transformations to tackle some of these hard problems with better-understood heuristics, and prove the tractability or fixed-parameter tractability of more modest queries which retain useful affordances. This framework allows us to understand the scope and limits of interpretability queries, explore viable options, and compare their resource demands on existing and future architectures.

We propose Scalable Mechanistic Neural Network (S-MNN), an enhanced neural network framework designed for scientific machine learning applications involving long temporal sequences. By reformulating the original Mechanistic Neural Network (MNN) (Pervez et al., 2024), we reduce the computational time and space complexities from cubic and quadratic with respect to the sequence length, respectively, to linear. This significant improvement enables efficient modeling of long-term dynamics without sacrificing accuracy or interpretability. Extensive experiments demonstrate that S-MNN matches the original MNN in precision while substantially reducing computational resources. Consequently, S-MNN can drop-in replace the original MNN in applications, providing a practical and efficient tool for integrating mechanistic bottlenecks into neural network models of complex dynamical systems. Source code is available at https://github.com/IST-DASLab/ScalableMNN.