Daily Papers

Language confusion -- where large language models (LLMs) generate unintended languages against the user's need -- remains a critical challenge, especially for English-centric models. We present the first mechanistic interpretability (MI) study of language confusion, combining behavioral benchmarking with neuron-level analysis. Using the Language Confusion Benchmark (LCB), we show that confusion points (CPs) -- specific positions where language switches occur -- are central to this phenomenon. Through layer-wise analysis with TunedLens and targeted neuron attribution, we reveal that transition failures in the final layers drive confusion. We further demonstrate that editing a small set of critical neurons, identified via comparative analysis with multilingual-tuned models, substantially mitigates confusion without harming general competence or fluency. Our approach matches multilingual alignment in confusion reduction for most languages and yields cleaner, higher-quality outputs. These findings provide new insights into the internal dynamics of LLMs and highlight neuron-level interventions as a promising direction for robust, interpretable multilingual language modeling.

While alignment algorithms are now commonly used to tune pre-trained language models towards a user's preferences, we lack explanations for the underlying mechanisms in which models become ``aligned'', thus making it difficult to explain phenomena like jailbreaks. In this work we study a popular algorithm, direct preference optimization (DPO), and the mechanisms by which it reduces toxicity. Namely, we first study how toxicity is represented and elicited in a pre-trained language model, GPT2-medium. We then apply DPO with a carefully crafted pairwise dataset to reduce toxicity. We examine how the resulting model averts toxic outputs, and find that capabilities learned from pre-training are not removed, but rather bypassed. We use this insight to demonstrate a simple method to un-align the model, reverting it back to its toxic behavior.

Despite superior reasoning prowess demonstrated by Large Language Models (LLMs) with Chain-of-Thought (CoT) prompting, a lack of understanding prevails around the internal mechanisms of the models that facilitate CoT generation. This work investigates the neural sub-structures within LLMs that manifest CoT reasoning from a mechanistic point of view. From an analysis of LLaMA-2 7B applied to multistep reasoning over fictional ontologies, we demonstrate that LLMs deploy multiple parallel pathways of answer generation for step-by-step reasoning. These parallel pathways provide sequential answers from the input question context as well as the generated CoT. We observe a striking functional rift in the middle layers of the LLM. Token representations in the initial half remain strongly biased towards the pretraining prior, with the in-context taking over abruptly in the later half. This internal phase shift manifests in different functional components: attention heads that write the answer token predominantly appear in the later half, attention heads that move information along ontological relationships appear exclusively in the initial half, and so on. To the best of our knowledge, this is the first attempt towards mechanistic investigation of CoT reasoning in LLMs.

While the successes of transformers across many domains are indisputable, accurate understanding of the learning mechanics is still largely lacking. Their capabilities have been probed on benchmarks which include a variety of structured and reasoning tasks -- but mathematical understanding is lagging substantially behind. Recent lines of work have begun studying representational aspects of this question: that is, the size/depth/complexity of attention-based networks to perform certain tasks. However, there is no guarantee the learning dynamics will converge to the constructions proposed. In our paper, we provide fine-grained mechanistic understanding of how transformers learn "semantic structure", understood as capturing co-occurrence structure of words. Precisely, we show, through a combination of experiments on synthetic data modeled by Latent Dirichlet Allocation (LDA), Wikipedia data, and mathematical analysis that the embedding layer and the self-attention layer encode the topical structure. In the former case, this manifests as higher average inner product of embeddings between same-topic words. In the latter, it manifests as higher average pairwise attention between same-topic words. The mathematical results involve several assumptions to make the analysis tractable, which we verify on data, and might be of independent interest as well.

Large language models are increasingly used to process documents and facilitate question-answering on them. In our paper, we extract mechanistic circuits for this real-world language modeling task: context-augmented language modeling for extractive question-answering (QA) tasks and understand the potential benefits of circuits towards downstream applications such as data attribution to context information. We extract circuits as a function of internal model components (e.g., attention heads, MLPs) using causal mediation analysis techniques. Leveraging the extracted circuits, we first understand the interplay between the model's usage of parametric memory and retrieved context towards a better mechanistic understanding of context-augmented language models. We then identify a small set of attention heads in our circuit which performs reliable data attribution by default, thereby obtaining attribution for free in just the model's forward pass. Using this insight, we then introduce ATTNATTRIB, a fast data attribution algorithm which obtains state-of-the-art attribution results across various extractive QA benchmarks. Finally, we show the possibility to steer the language model towards answering from the context, instead of the parametric memory by using the attribution from ATTNATTRIB as an additional signal during the forward pass. Beyond mechanistic understanding, our paper provides tangible applications of circuits in the form of reliable data attribution and model steering.

A mechanistic understanding of how MLPs do computation in deep neural networks 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 nevertheless 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 eigendecomposition reveals interpretable low-rank structure across toy tasks, image classification, 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.

In the field of antibody engineering, an essential task is to design a novel antibody whose paratopes bind to a specific antigen with correct epitopes. Understanding antibody structure and its paratope can facilitate a mechanistic understanding of its function. Therefore, antibody structure prediction from its sequence alone has always been a highly valuable problem for de novo antibody design. AlphaFold2, a breakthrough in the field of structural biology, provides a solution to predict protein structure based on protein sequences and computationally expensive coevolutionary multiple sequence alignments (MSAs). However, the computational efficiency and undesirable prediction accuracy of antibodies, especially on the complementarity-determining regions (CDRs) of antibodies limit their applications in the industrially high-throughput drug design. To learn an informative representation of antibodies, we employed a deep antibody language model (ALM) on curated sequences from the observed antibody space database via a transformer model. We also developed a novel model named xTrimoABFold to predict antibody structure from antibody sequence based on the pretrained ALM as well as efficient evoformers and structural modules. The model was trained end-to-end on the antibody structures in PDB by minimizing the ensemble loss of domain-specific focal loss on CDR and the frame-aligned point loss. xTrimoABFold outperforms AlphaFold2 and other protein language model based SOTAs, e.g., OmegaFold, HelixFold-Single, and IgFold with a large significant margin (30+\% improvement on RMSD) while performing 151 times faster than AlphaFold2. To the best of our knowledge, xTrimoABFold achieved state-of-the-art antibody structure prediction. Its improvement in both accuracy and efficiency makes it a valuable tool for de novo antibody design and could make further improvements in immuno-theory.

Research in mechanistic interpretability seeks to explain behaviors of machine learning models in terms of their internal components. However, most previous work either focuses on simple behaviors in small models, or describes complicated behaviors in larger models with broad strokes. In this work, we bridge this gap by presenting an explanation for how GPT-2 small performs a natural language task called indirect object identification (IOI). Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior "in the wild" in a language model. We evaluate the reliability of our explanation using three quantitative criteria--faithfulness, completeness and minimality. Though these criteria support our explanation, they also point to remaining gaps in our understanding. Our work provides evidence that a mechanistic understanding of large ML models is feasible, opening opportunities to scale our understanding to both larger models and more complex tasks.

We investigate the internal behavior of Transformer-based Large Language Models (LLMs) when they generate factually incorrect text. We propose modeling factual queries as Constraint Satisfaction Problems and use this framework to investigate how the model interacts internally with factual constraints. Specifically, we discover a strong positive relation between the model's attention to constraint tokens and the factual accuracy of its responses. In our curated suite of 11 datasets with over 40,000 prompts, we study the task of predicting factual errors with the Llama-2 family across all scales (7B, 13B, 70B). We propose SAT Probe, a method probing self-attention patterns, that can predict constraint satisfaction and factual errors, and allows early error identification. The approach and findings demonstrate how using the mechanistic understanding of factuality in LLMs can enhance reliability.

Interpretability provides a toolset for understanding how and why neural networks behave in certain ways. However, there is little unity in the field: most studies employ ad-hoc evaluations and do not share theoretical foundations, making it difficult to measure progress and compare the pros and cons of different techniques. Furthermore, while mechanistic understanding is frequently discussed, the basic causal units underlying these mechanisms are often not explicitly defined. In this paper, we propose a perspective on interpretability research grounded in causal mediation analysis. Specifically, we describe the history and current state of interpretability taxonomized according to the types of causal units (mediators) employed, as well as methods used to search over mediators. We discuss the pros and cons of each mediator, providing insights as to when particular kinds of mediators and search methods are most appropriate depending on the goals of a given study. We argue that this framing yields a more cohesive narrative of the field, as well as actionable insights for future work. Specifically, we recommend a focus on discovering new mediators with better trade-offs between human-interpretability and compute-efficiency, and which can uncover more sophisticated abstractions from neural networks than the primarily linear mediators employed in current work. We also argue for more standardized evaluations that enable principled comparisons across mediator types, such that we can better understand when particular causal units are better suited to particular use cases.

Large language models (LLMs) excel on a variety of reasoning benchmarks, but previous studies suggest they sometimes struggle to generalize to unseen questions, potentially due to over-reliance on memorized training examples. However, the precise conditions under which LLMs switch between reasoning and memorization during text generation remain unclear. In this work, we provide a mechanistic understanding of LLMs' reasoning-memorization dynamics by identifying a set of linear features in the model's residual stream that govern the balance between genuine reasoning and memory recall. These features not only distinguish reasoning tasks from memory-intensive ones but can also be manipulated to causally influence model performance on reasoning tasks. Additionally, we show that intervening in these reasoning features helps the model more accurately activate the most relevant problem-solving capabilities during answer generation. Our findings offer new insights into the underlying mechanisms of reasoning and memory in LLMs and pave the way for the development of more robust and interpretable generative AI systems.

Refusal mechanisms in large language models (LLMs) are essential for ensuring safety. Recent research has revealed that refusal behavior can be mediated by a single direction in activation space, enabling targeted interventions to bypass refusals. While this is primarily demonstrated in an English-centric context, appropriate refusal behavior is important for any language, but poorly understood. In this paper, we investigate the refusal behavior in LLMs across 14 languages using PolyRefuse, a multilingual safety dataset created by translating malicious and benign English prompts into these languages. We uncover the surprising cross-lingual universality of the refusal direction: a vector extracted from English can bypass refusals in other languages with near-perfect effectiveness, without any additional fine-tuning. Even more remarkably, refusal directions derived from any safety-aligned language transfer seamlessly to others. We attribute this transferability to the parallelism of refusal vectors across languages in the embedding space and identify the underlying mechanism behind cross-lingual jailbreaks. These findings provide actionable insights for building more robust multilingual safety defenses and pave the way for a deeper mechanistic understanding of cross-lingual vulnerabilities in LLMs.

Central to our understanding of chemical reactivity is the principle of mass conservation, which is fundamental for ensuring physical consistency, balancing equations, and guiding reaction design. However, data-driven computational models for tasks such as reaction product prediction rarely abide by this most basic constraint. In this work, we recast the problem of reaction prediction as a problem of electron redistribution using the modern deep generative framework of flow matching. Our model, FlowER, overcomes limitations inherent in previous approaches by enforcing exact mass conservation, thereby resolving hallucinatory failure modes, recovering mechanistic reaction sequences for unseen substrate scaffolds, and generalizing effectively to out-of-domain reaction classes with extremely data-efficient fine-tuning. FlowER additionally enables estimation of thermodynamic or kinetic feasibility and manifests a degree of chemical intuition in reaction prediction tasks. This inherently interpretable framework represents a significant step in bridging the gap between predictive accuracy and mechanistic understanding in data-driven reaction outcome prediction.

Despite the growing use of transformer models in computer vision, a mechanistic understanding of these networks is still needed. This work introduces a method to reverse-engineer Vision Transformers trained to solve image classification tasks. Inspired by previous research in NLP, we demonstrate how the inner representations at any level of the hierarchy can be projected onto the learned class embedding space to uncover how these networks build categorical representations for their predictions. We use our framework to show how image tokens develop class-specific representations that depend on attention mechanisms and contextual information, and give insights on how self-attention and MLP layers differentially contribute to this categorical composition. We additionally demonstrate that this method (1) can be used to determine the parts of an image that would be important for detecting the class of interest, and (2) exhibits significant advantages over traditional linear probing approaches. Taken together, our results position our proposed framework as a powerful tool for mechanistic interpretability and explainability research.

Language models learn a great quantity of factual information during pretraining, and recent work localizes this information to specific model weights like mid-layer MLP weights. In this paper, we find that we can change how a fact is stored in a model by editing weights that are in a different location than where existing methods suggest that the fact is stored. This is surprising because we would expect that localizing facts to specific model parameters would tell us where to manipulate knowledge in models, and this assumption has motivated past work on model editing methods. Specifically, we show that localization conclusions from representation denoising (also known as Causal Tracing) do not provide any insight into which model MLP layer would be best to edit in order to override an existing stored fact with a new one. This finding raises questions about how past work relies on Causal Tracing to select which model layers to edit. Next, we consider several variants of the editing problem, including erasing and amplifying facts. For one of our editing problems, editing performance does relate to localization results from representation denoising, but we find that which layer we edit is a far better predictor of performance. Our results suggest, counterintuitively, that better mechanistic understanding of how pretrained language models work may not always translate to insights about how to best change their behavior. Our code is available at https://github.com/google/belief-localization

This paper investigates the ability of transformer-based models to learn structural recursion from examples. Recursion is a universal concept in both natural and formal languages. Structural recursion is central to the programming language and formal mathematics tasks where symbolic tools currently excel beyond neural models, such as inferring semantic relations between datatypes and emulating program behavior. We introduce a general framework that nicely connects the abstract concepts of structural recursion in the programming language domain to concrete sequence modeling problems and learned models' behavior. The framework includes a representation that captures the general syntax of structural recursion, coupled with two different frameworks for understanding their semantics -- one that is more natural from a programming languages perspective and one that helps bridge that perspective with a mechanistic understanding of the underlying transformer architecture. With our framework as a powerful conceptual tool, we identify different issues under various set-ups. The models trained to emulate recursive computations cannot fully capture the recursion yet instead fit short-cut algorithms and thus cannot solve certain edge cases that are under-represented in the training distribution. In addition, it is difficult for state-of-the-art large language models (LLMs) to mine recursive rules from in-context demonstrations. Meanwhile, these LLMs fail in interesting ways when emulating reduction (step-wise computation) of the recursive function.

Fine-tuning adapts pretrained models for specific tasks but poses the risk of catastrophic forgetting (CF), where critical knowledge from pretraining is overwritten. To address the issue of CF in a general-purpose framework, we propose Low-damage Knowledge Implanting (LoKI), a parameter-efficient fine-tuning (PEFT) technique that utilizes recent mechanistic understanding of how knowledge is stored in transformer architectures. We compare LoKI against state-of-the-art PEFT methods in two real-world fine-tuning scenarios. The results show that LoKI demonstrates significantly better preservation of general capabilities. At the same time, its task-specific performance is comparable to or even surpasses that of full parameter fine-tuning and these PEFT methods across various model architectures. Our work bridges the mechanistic insights of LLMs' knowledge storage with practical fine-tuning objectives, enabling an effective balance between task-specific adaptation and the retention of general-purpose capabilities.

While large language models based on the transformer architecture have demonstrated remarkable in-context learning (ICL) capabilities, understandings of such capabilities are still in an early stage, where existing theory and mechanistic understanding focus mostly on simple scenarios such as learning simple function classes. This paper takes initial steps on understanding ICL in more complex scenarios, by studying learning with representations. Concretely, we construct synthetic in-context learning problems with a compositional structure, where the label depends on the input through a possibly complex but fixed representation function, composed with a linear function that differs in each instance. By construction, the optimal ICL algorithm first transforms the inputs by the representation function, and then performs linear ICL on top of the transformed dataset. We show theoretically the existence of transformers that approximately implement such algorithms with mild depth and size. Empirically, we find trained transformers consistently achieve near-optimal ICL performance in this setting, and exhibit the desired dissection where lower layers transforms the dataset and upper layers perform linear ICL. Through extensive probing and a new pasting experiment, we further reveal several mechanisms within the trained transformers, such as concrete copying behaviors on both the inputs and the representations, linear ICL capability of the upper layers alone, and a post-ICL representation selection mechanism in a harder mixture setting. These observed mechanisms align well with our theory and may shed light on how transformers perform ICL in more realistic scenarios.

Interpretability can improve the safety, transparency and trust of AI models, which is especially important in healthcare applications where decisions often carry significant consequences. Mechanistic interpretability, particularly through the use of sparse autoencoders (SAEs), offers a promising approach for uncovering human-interpretable features within large transformer-based models. In this study, we apply Matryoshka-SAE to the radiology-specialised multimodal large language model, MAIRA-2, to interpret its internal representations. Using large-scale automated interpretability of the SAE features, we identify a range of clinically relevant concepts - including medical devices (e.g., line and tube placements, pacemaker presence), pathologies such as pleural effusion and cardiomegaly, longitudinal changes and textual features. We further examine the influence of these features on model behaviour through steering, demonstrating directional control over generations with mixed success. Our results reveal practical and methodological challenges, yet they offer initial insights into the internal concepts learned by MAIRA-2 - marking a step toward deeper mechanistic understanding and interpretability of a radiology-adapted multimodal large language model, and paving the way for improved model transparency. We release the trained SAEs and interpretations: https://huggingface.co/microsoft/maira-2-sae.

The transcriptional response to genetic perturbation reveals fundamental insights into complex cellular systems. While current approaches have made progress in predicting genetic perturbation responses, they provide limited biological understanding and cannot systematically refine existing knowledge. Overcoming these limitations requires an end-to-end integration of data-driven learning and existing knowledge. However, this integration is challenging due to inconsistencies between data and knowledge bases, such as noise, misannotation, and incompleteness. To address this challenge, we propose ALIGNED (Adaptive aLignment for Inconsistent Genetic kNowledgE and Data), a neuro-symbolic framework based on the Abductive Learning (ABL) paradigm. This end-to-end framework aligns neural and symbolic components and performs systematic knowledge refinement. We introduce a balanced consistency metric to evaluate the predictions' consistency against both data and knowledge. Our results show that ALIGNED outperforms state-of-the-art methods by achieving the highest balanced consistency, while also re-discovering biologically meaningful knowledge. Our work advances beyond existing methods to enable both the transparency and the evolution of mechanistic biological understanding.

Predicting high-dimensional or extreme multilabels, such as in medical coding, requires both accuracy and interpretability. Existing works often rely on local interpretability methods, failing to provide comprehensive explanations of the overall mechanism behind each label prediction within a multilabel set. We propose a mechanistic interpretability module called DIctionary Label Attention (\method) that disentangles uninterpretable dense embeddings into a sparse embedding space, where each nonzero element (a dictionary feature) represents a globally learned medical concept. Through human evaluations, we show that our sparse embeddings are more human understandable than its dense counterparts by at least 50 percent. Our automated dictionary feature identification pipeline, leveraging large language models (LLMs), uncovers thousands of learned medical concepts by examining and summarizing the highest activating tokens for each dictionary feature. We represent the relationships between dictionary features and medical codes through a sparse interpretable matrix, enhancing the mechanistic and global understanding of the model's predictions while maintaining competitive performance and scalability without extensive human annotation.

The ability of neural networks to represent more features than neurons makes interpreting them challenging. This phenomenon, known as superposition, has spurred efforts to find architectures that are more interpretable than standard multilayer perceptrons (MLPs) with elementwise activation functions. In this note, I examine bilinear layers, which are a type of MLP layer that are mathematically much easier to analyze while simultaneously performing better than standard MLPs. Although they are nonlinear functions of their input, I demonstrate that bilinear layers can be expressed using only linear operations and third order tensors. We can integrate this expression for bilinear layers into a mathematical framework for transformer circuits, which was previously limited to attention-only transformers. These results suggest that bilinear layers are easier to analyze mathematically than current architectures and thus may lend themselves to deeper safety insights by allowing us to talk more formally about circuits in neural networks. Additionally, bilinear layers may offer an alternative path for mechanistic interpretability through understanding the mechanisms of feature construction instead of enumerating a (potentially exponentially) large number of features in large models.

Spider silks are remarkable materials characterized by superb mechanical properties such as strength, extensibility and lightweightedness. Yet, to date, limited models are available to fully explore sequence-property relationships for analysis and design. Here we propose a custom generative large-language model to enable design of novel spider silk protein sequences to meet complex combinations of target mechanical properties. The model, pretrained on a large set of protein sequences, is fine-tuned on ~1,000 major ampullate spidroin (MaSp) sequences for which associated fiber-level mechanical properties exist, to yield an end-to-end forward and inverse generative strategy. Performance is assessed through: (1), a novelty analysis and protein type classification for generated spidroin sequences through BLAST searches, (2) property evaluation and comparison with similar sequences, (3) comparison of molecular structures, as well as, and (4) a detailed sequence motif analyses. We generate silk sequences with property combinations that do not exist in nature, and develop a deep understanding the mechanistic roles of sequence patterns in achieving overarching key mechanical properties (elastic modulus, strength, toughness, failure strain). The model provides an efficient approach to expand the silkome dataset, facilitating further sequence-structure analyses of silks, and establishes a foundation for synthetic silk design and optimization.

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: https://github.com/zepingyu0512/llava-mechanism

Understanding how features evolve across layers in deep neural networks is a fundamental challenge in mechanistic interpretability, particularly due to polysemanticity and feature superposition. While Sparse Autoencoders (SAEs) have been used to extract interpretable features from individual layers, aligning these features across layers has remained an open problem. In this paper, we introduce SAE Match, a novel, data-free method for aligning SAE features across different layers of a neural network. Our approach involves matching features by minimizing the mean squared error between the folded parameters of SAEs, a technique that incorporates activation thresholds into the encoder and decoder weights to account for differences in feature scales. Through extensive experiments on the Gemma 2 language model, we demonstrate that our method effectively captures feature evolution across layers, improving feature matching quality. We also show that features persist over several layers and that our approach can approximate hidden states across layers. Our work advances the understanding of feature dynamics in neural networks and provides a new tool for mechanistic interpretability studies.

The integration of Rotary Position Embedding (RoPE) in Multimodal Diffusion Transformer (MMDiT) has significantly enhanced text-to-image generation quality. However, the fundamental reliance of self-attention layers on positional embedding versus query-key similarity during generation remains an intriguing question. We present the first mechanistic analysis of RoPE-based MMDiT models (e.g., FLUX), introducing an automated probing strategy that disentangles positional information versus content dependencies by strategically manipulating RoPE during generation. Our analysis reveals distinct dependency patterns that do not straightforwardly correlate with depth, offering new insights into the layer-specific roles in RoPE-based MMDiT. Based on these findings, we propose a training-free, task-specific image editing framework that categorizes editing tasks into three types: position-dependent editing (e.g., object addition), content similarity-dependent editing (e.g., non-rigid editing), and region-preserved editing (e.g., background replacement). For each type, we design tailored key-value injection strategies based on the characteristics of the editing task. Extensive qualitative and quantitative evaluations demonstrate that our method outperforms state-of-the-art approaches, particularly in preserving original semantic content and achieving seamless modifications.

Large Language Models (LLMs) have demonstrated unprecedented capabilities across various natural language processing tasks. Their ability to process and generate viable text and code has made them ubiquitous in many fields, while their deployment as knowledge bases and "reasoning" tools remains an area of ongoing research. In geography, a growing body of literature has been focusing on evaluating LLMs' geographical knowledge and their ability to perform spatial reasoning. However, very little is still known about the internal functioning of these models, especially about how they process geographical information. In this chapter, we establish a novel framework for the study of geospatial mechanistic interpretability - using spatial analysis to reverse engineer how LLMs handle geographical information. Our aim is to advance our understanding of the internal representations that these complex models generate while processing geographical information - what one might call "how LLMs think about geographic information" if such phrasing was not an undue anthropomorphism. We first outline the use of probing in revealing internal structures within LLMs. We then introduce the field of mechanistic interpretability, discussing the superposition hypothesis and the role of sparse autoencoders in disentangling polysemantic internal representations of LLMs into more interpretable, monosemantic features. In our experiments, we use spatial autocorrelation to show how features obtained for placenames display spatial patterns related to their geographic location and can thus be interpreted geospatially, providing insights into how these models process geographical information. We conclude by discussing how our framework can help shape the study and use of foundation models in geography.

Transformers demonstrate impressive performance on a range of reasoning benchmarks. To evaluate the degree to which these abilities are a result of actual reasoning, existing work has focused on developing sophisticated benchmarks for behavioral studies. However, these studies do not provide insights into the internal mechanisms driving the observed capabilities. To improve our understanding of the internal mechanisms of transformers, we present a comprehensive mechanistic analysis of a transformer trained on a synthetic reasoning task. We identify a set of interpretable mechanisms the model uses to solve the task, and validate our findings using correlational and causal evidence. Our results suggest that it implements a depth-bounded recurrent mechanisms that operates in parallel and stores intermediate results in selected token positions. We anticipate that the motifs we identified in our synthetic setting can provide valuable insights into the broader operating principles of transformers and thus provide a basis for understanding more complex models.

Mathematical reasoning in large language models (LMs) has garnered significant attention in recent work, but there is a limited understanding of how these models process and store information related to arithmetic tasks within their architecture. In order to improve our understanding of this aspect of language models, we present a mechanistic interpretation of Transformer-based LMs on arithmetic questions using a causal mediation analysis framework. By intervening on the activations of specific model components and measuring the resulting changes in predicted probabilities, we identify the subset of parameters responsible for specific predictions. This provides insights into how information related to arithmetic is processed by LMs. Our experimental results indicate that LMs process the input by transmitting the information relevant to the query from mid-sequence early layers to the final token using the attention mechanism. Then, this information is processed by a set of MLP modules, which generate result-related information that is incorporated into the residual stream. To assess the specificity of the observed activation dynamics, we compare the effects of different model components on arithmetic queries with other tasks, including number retrieval from prompts and factual knowledge questions.

Neural networks often exhibit emergent behavior, where qualitatively new capabilities arise from scaling up the amount of parameters, training data, or training steps. One approach to understanding emergence is to find continuous progress measures that underlie the seemingly discontinuous qualitative changes. We argue that progress measures can be found via mechanistic interpretability: reverse-engineering learned behaviors into their individual components. As a case study, we investigate the recently-discovered phenomenon of ``grokking'' exhibited by small transformers trained on modular addition tasks. We fully reverse engineer the algorithm learned by these networks, which uses discrete Fourier transforms and trigonometric identities to convert addition to rotation about a circle. We confirm the algorithm by analyzing the activations and weights and by performing ablations in Fourier space. Based on this understanding, we define progress measures that allow us to study the dynamics of training and split training into three continuous phases: memorization, circuit formation, and cleanup. Our results show that grokking, rather than being a sudden shift, arises from the gradual amplification of structured mechanisms encoded in the weights, followed by the later removal of memorizing components.

Text-to-image (T2I) diffusion models rely on encoded prompts to guide the image generation process. Typically, these prompts are extended to a fixed length by adding padding tokens before text encoding. Despite being a default practice, the influence of padding tokens on the image generation process has not been investigated. In this work, we conduct the first in-depth analysis of the role padding tokens play in T2I models. We develop two causal techniques to analyze how information is encoded in the representation of tokens across different components of the T2I pipeline. Using these techniques, we investigate when and how padding tokens impact the image generation process. Our findings reveal three distinct scenarios: padding tokens may affect the model's output during text encoding, during the diffusion process, or be effectively ignored. Moreover, we identify key relationships between these scenarios and the model's architecture (cross or self-attention) and its training process (frozen or trained text encoder). These insights contribute to a deeper understanding of the mechanisms of padding tokens, potentially informing future model design and training practices in T2I systems.

Large neural models are increasingly deployed in high-stakes settings, raising concerns about whether their behavior reliably aligns with human values. Interpretability provides a route to internal transparency by revealing the computations that drive outputs. We argue that interpretability especially mechanistic approaches should be treated as a design principle for alignment, not an auxiliary diagnostic tool. Post-hoc methods such as LIME or SHAP offer intuitive but correlational explanations, while mechanistic techniques like circuit tracing or activation patching yield causal insight into internal failures, including deceptive or misaligned reasoning that behavioral methods like RLHF, red teaming, or Constitutional AI may overlook. Despite these advantages, interpretability faces challenges of scalability, epistemic uncertainty, and mismatches between learned representations and human concepts. Our position is that progress on safe and trustworthy AI will depend on making interpretability a first-class objective of AI research and development, ensuring that systems are not only effective but also auditable, transparent, and aligned with human intent.

Understanding training dynamics and feature evolution is crucial for the mechanistic interpretability of large language models (LLMs). Although sparse autoencoders (SAEs) have been used to identify features within LLMs, a clear picture of how these features evolve during training remains elusive. In this study, we: (1) introduce SAE-Track, a novel method to efficiently obtain a continual series of SAEs; (2) mechanistically investigate feature formation and develop a progress measure for it ; and (3) analyze and visualize feature drift during training. Our work provides new insights into the dynamics of features in LLMs, enhancing our understanding of training mechanisms and feature evolution.

Transformers exhibit impressive capabilities but are often regarded as black boxes due to challenges in understanding the complex nonlinear relationships between features. Interpreting machine learning models is of paramount importance to mitigate risks, and mechanistic interpretability is in particular of current interest as it opens up a window for guiding manual modifications and reverse-engineering solutions. In this work, we introduce contextual decomposition for transformers (CD-T), extending a prior work on CD for RNNs and CNNs, to address mechanistic interpretation computationally efficiently. CD-T is a flexible interpretation method for transformers. It can capture contributions of combinations of input features or source internal components (e.g. attention heads, feed-forward networks) to (1) final predictions or (2) the output of any target internal component. Using CD-T, we propose a novel algorithm for circuit discovery. On a real-world pathology report classification task: we show CD-T distills a more faithful circuit of attention heads with improved computational efficiency (speed up 2x) than a prior benchmark, path patching. As a versatile interpretation method, CD-T also exhibits exceptional capabilities for local interpretations. CD-T is shown to reliably find words and phrases of contrasting sentiment/topic on SST-2 and AGNews datasets. Through human experiments, we demonstrate CD-T enables users to identify the more accurate of two models and to better trust a model's outputs compared to alternative interpretation methods such as SHAP and LIME.

Robust tooling and publicly available pre-trained models have helped drive recent advances in mechanistic interpretability for language models. However, similar progress in vision mechanistic interpretability has been hindered by the lack of accessible frameworks and pre-trained weights. We present Prisma (Access the codebase here: https://github.com/Prisma-Multimodal/ViT-Prisma), an open-source framework designed to accelerate vision mechanistic interpretability research, providing a unified toolkit for accessing 75+ vision and video transformers; support for sparse autoencoder (SAE), transcoder, and crosscoder training; a suite of 80+ pre-trained SAE weights; activation caching, circuit analysis tools, and visualization tools; and educational resources. Our analysis reveals surprising findings, including that effective vision SAEs can exhibit substantially lower sparsity patterns than language SAEs, and that in some instances, SAE reconstructions can decrease model loss. Prisma enables new research directions for understanding vision model internals while lowering barriers to entry in this emerging field.

Training recurrent neural networks (RNNs) is a high-dimensional process that requires updating numerous parameters. Therefore, it is often difficult to pinpoint the underlying learning mechanisms. To address this challenge, we propose to gain mechanistic insights into the phenomenon of abrupt learning by studying RNNs trained to perform diverse short-term memory tasks. In these tasks, RNN training begins with an initial search phase. Following a long period of plateau in accuracy, the values of the loss function suddenly drop, indicating abrupt learning. Analyzing the neural computation performed by these RNNs reveals geometric restructuring (GR) in their phase spaces prior to the drop. To promote these GR events, we introduce a temporal consistency regularization that accelerates (bioplausible) training, facilitates attractor formation, and enables efficient learning in strongly connected networks. Our findings offer testable predictions for neuroscientists and emphasize the need for goal-agnostic secondary mechanisms to facilitate learning in biological and artificial networks.

Large language models (LLMs) have made remarkable progress in various domains, yet they often suffer from repetitive text generation, a phenomenon we refer to as the "Repeat Curse". While previous studies have proposed decoding strategies to mitigate repetition, the underlying mechanism behind this issue remains insufficiently explored. In this work, we investigate the root causes of repetition in LLMs through the lens of mechanistic interpretability. Inspired by recent advances in Sparse Autoencoders (SAEs), which enable monosemantic feature extraction, we propose a novel approach, "Duplicatus Charm", to induce and analyze the Repeat Curse. Our method systematically identifies "Repetition Features" -the key model activations responsible for generating repetitive outputs. First, we locate the layers most involved in repetition through logit analysis. Next, we extract and stimulate relevant features using SAE-based activation manipulation. To validate our approach, we construct a repetition dataset covering token and paragraph level repetitions and introduce an evaluation pipeline to quantify the influence of identified repetition features. Furthermore, by deactivating these features, we have effectively mitigated the Repeat Curse.

Compositional relational reasoning (CRR) is a hallmark of human intelligence, but we lack a clear understanding of whether and how existing transformer large language models (LLMs) can solve CRR tasks. To enable systematic exploration of the CRR capability of LLMs, we first propose a new synthetic benchmark called Generalized Associative Recall (GAR) by integrating and generalizing the essence of several tasks in mechanistic interpretability (MI) study in a unified framework. Evaluation shows that GAR is challenging enough for existing LLMs, revealing their fundamental deficiency in CRR. Meanwhile, it is easy enough for systematic MI study. Then, to understand how LLMs solve GAR tasks, we use attribution patching to discover the core circuits reused by Vicuna-33B across different tasks and a set of vital attention heads. Intervention experiments show that the correct functioning of these heads significantly impacts task performance. Especially, we identify two classes of heads whose activations represent the abstract notion of true and false in GAR tasks respectively. They play a fundamental role in CRR across various models and tasks. The dataset and code are available at https://github.com/Caiyun-AI/GAR.

Stepwise inference protocols, such as scratchpads and chain-of-thought, help language models solve complex problems by decomposing them into a sequence of simpler subproblems. Despite the significant gain in performance achieved via these protocols, the underlying mechanisms of stepwise inference have remained elusive. To address this, we propose to study autoregressive Transformer models on a synthetic task that embodies the multi-step nature of problems where stepwise inference is generally most useful. Specifically, we define a graph navigation problem wherein a model is tasked with traversing a path from a start to a goal node on the graph. Despite is simplicity, we find we can empirically reproduce and analyze several phenomena observed at scale: (i) the stepwise inference reasoning gap, the cause of which we find in the structure of the training data; (ii) a diversity-accuracy tradeoff in model generations as sampling temperature varies; (iii) a simplicity bias in the model's output; and (iv) compositional generalization and a primacy bias with in-context exemplars. Overall, our work introduces a grounded, synthetic framework for studying stepwise inference and offers mechanistic hypotheses that can lay the foundation for a deeper understanding of this phenomenon.

Modern language models can imitate complex patterns through few-shot learning, enabling them to complete challenging tasks without fine-tuning. However, imitation can also lead models to reproduce inaccuracies or harmful content if present in the context. We study harmful imitation through the lens of a model's internal representations, and identify two related phenomena: "overthinking" and "false induction heads". The first phenomenon, overthinking, appears when we decode predictions from intermediate layers, given correct vs. incorrect few-shot demonstrations. At early layers, both demonstrations induce similar model behavior, but the behavior diverges sharply at some "critical layer", after which the accuracy given incorrect demonstrations progressively decreases. The second phenomenon, false induction heads, are a possible mechanistic cause of overthinking: these are heads in late layers that attend to and copy false information from previous demonstrations, and whose ablation reduces overthinking. Beyond scientific understanding, our results suggest that studying intermediate model computations could be a promising avenue for understanding and guarding against harmful model behaviors.

The mechanisms by which reasoning training reshapes LLMs' internal computations remain unclear. We study lightweight steering vectors inserted into the base model's residual stream and trained with a reinforcement-learning objective. These vectors match full fine-tuning performance while preserving the interpretability of small, additive interventions. Using logit-lens readouts and path-patching analyses on two models, we find that (i) the last-layer steering vector acts like a token-substitution bias concentrated on the first generated token, consistently boosting tokens such as "To" and "Step"; (ii) the penultimate-layer vector leaves attention patterns largely intact and instead operates through the MLP and unembedding, preferentially up-weighting process words and structure symbols; and (iii) middle layers de-emphasize non-English tokens. Next, we show that a SAE isolates features associated with correct generations. We also show that steering vectors (i) transfer to other models, (ii) combine across layers when trained in isolation, and (iii) concentrate magnitude on meaningful prompt segments under adaptive token-wise scaling. Taken together, these results deepen understanding of how trained steering vectors shape computation and should inform future work in activation engineering and the study of reasoning models.

We study how multi-head softmax attention models are trained to perform in-context learning on linear data. Through extensive empirical experiments and rigorous theoretical analysis, we demystify the emergence of elegant attention patterns: a diagonal and homogeneous pattern in the key-query (KQ) weights, and a last-entry-only and zero-sum pattern in the output-value (OV) weights. Remarkably, these patterns consistently appear from gradient-based training starting from random initialization. Our analysis reveals that such emergent structures enable multi-head attention to approximately implement a debiased gradient descent predictor -- one that outperforms single-head attention and nearly achieves Bayesian optimality up to proportional factor. Furthermore, compared to linear transformers, the softmax attention readily generalizes to sequences longer than those seen during training. We also extend our study to scenarios with non-isotropic covariates and multi-task linear regression. In the former, multi-head attention learns to implement a form of pre-conditioned gradient descent. In the latter, we uncover an intriguing regime where the interplay between head number and task number triggers a superposition phenomenon that efficiently resolves multi-task in-context learning. Our results reveal that in-context learning ability emerges from the trained transformer as an aggregated effect of its architecture and the underlying data distribution, paving the way for deeper understanding and broader applications of in-context learning.

Layer-wise normalization (LN) is an essential component of virtually all transformer-based large language models. While its effects on training stability are well documented, its role at inference time is poorly understood. Additionally, LN layers hinder mechanistic interpretability by introducing additional nonlinearities and increasing the interconnectedness of individual model components. Here, we show that all LN layers can be removed from every GPT-2 model with only a small increase in validation loss (e.g. +0.03 cross-entropy loss for GPT-2 XL). Thus, LN cannot play a substantial role in language modeling. We find that the amount of fine-tuning data needed for LN removal grows sublinearly with model parameters, suggesting scaling to larger models is feasible. We release a suite of LN-free GPT-2 models on Hugging Face. Furthermore, we test interpretability techniques on LN-free models. Direct logit attribution now gives the exact direct effect of individual components, while the accuracy of attribution patching does not significantly improve. We also confirm that GPT-2's "confidence neurons" are inactive in the LN-free models. Our work clarifies the role of LN layers in language modeling, showing that GPT-2-class models can function without LN layers. We hope that our LN-free analogs of the GPT-2 family of models will enable more precise interpretability research and improve our understanding of language models.

In-context learning is a powerful emergent ability in transformer models. Prior work in mechanistic interpretability has identified a circuit element that may be critical for in-context learning -- the induction head (IH), which performs a match-and-copy operation. During training of large transformers on natural language data, IHs emerge around the same time as a notable phase change in the loss. Despite the robust evidence for IHs and this interesting coincidence with the phase change, relatively little is known about the diversity and emergence dynamics of IHs. Why is there more than one IH, and how are they dependent on each other? Why do IHs appear all of a sudden, and what are the subcircuits that enable them to emerge? We answer these questions by studying IH emergence dynamics in a controlled setting by training on synthetic data. In doing so, we develop and share a novel optogenetics-inspired causal framework for modifying activations throughout training. Using this framework, we delineate the diverse and additive nature of IHs. By clamping subsets of activations throughout training, we then identify three underlying subcircuits that interact to drive IH formation, yielding the phase change. Furthermore, these subcircuits shed light on data-dependent properties of formation, such as phase change timing, already showing the promise of this more in-depth understanding of subcircuits that need to "go right" for an induction head.

Mechanistic Interpretability (MI) promises a path toward fully understanding how neural networks make their predictions. Prior work demonstrates that even when trained to perform simple arithmetic, models can implement a variety of algorithms (sometimes concurrently) depending on initialization and hyperparameters. Does this mean neuron-level interpretability techniques have limited applicability? We argue that high-dimensional neural networks can learn low-dimensional representations of their training data that are useful beyond simply making good predictions. Such representations can be understood through the mechanistic interpretability lens and provide insights that are surprisingly faithful to human-derived domain knowledge. This indicates that such approaches to interpretability can be useful for deriving a new understanding of a problem from models trained to solve it. As a case study, we extract nuclear physics concepts by studying models trained to reproduce nuclear data.

Understanding the decision-making processes of neural networks is a central goal of mechanistic interpretability. In the context of Large Language Models (LLMs), this involves uncovering the underlying mechanisms and identifying the roles of individual model components such as neurons and attention heads, as well as model abstractions such as the learned sparse features extracted by Sparse Autoencoders (SAEs). A rapidly growing line of work tackles this challenge by using powerful generator models to produce open-vocabulary, natural language concept descriptions for these components. In this paper, we provide the first survey of the emerging field of concept descriptions for model components and abstractions. We chart the key methods for generating these descriptions, the evolving landscape of automated and human metrics for evaluating them, and the datasets that underpin this research. Our synthesis reveals a growing demand for more rigorous, causal evaluation. By outlining the state of the art and identifying key challenges, this survey provides a roadmap for future research toward making models more transparent.

The greatest ambition of mechanistic interpretability is to completely rewrite deep neural networks in a format that is more amenable to human understanding, while preserving their behavior and performance. In this paper, we attempt to partially rewrite a large language model using simple natural language explanations. We first approximate one of the feedforward networks in the LLM with a wider MLP with sparsely activating neurons - a transcoder - and use an automated interpretability pipeline to generate explanations for these neurons. We then replace the first layer of this sparse MLP with an LLM-based simulator, which predicts the activation of each neuron given its explanation and the surrounding context. Finally, we measure the degree to which these modifications distort the model's final output. With our pipeline, the model's increase in loss is statistically similar to entirely replacing the sparse MLP output with the zero vector. We employ the same protocol, this time using a sparse autoencoder, on the residual stream of the same layer and obtain similar results. These results suggest that more detailed explanations are needed to improve performance substantially above the zero ablation baseline.

Large language models (LLMs) trained on huge corpora of text datasets demonstrate intriguing capabilities, achieving state-of-the-art performance on tasks they were not explicitly trained for. The precise nature of LLM capabilities is often mysterious, and different prompts can elicit different capabilities through in-context learning. We propose a framework that enables us to analyze in-context learning dynamics to understand latent concepts underlying LLMs' behavioral patterns. This provides a more nuanced understanding than success-or-failure evaluation benchmarks, but does not require observing internal activations as a mechanistic interpretation of circuits would. Inspired by the cognitive science of human randomness perception, we use random binary sequences as context and study dynamics of in-context learning by manipulating properties of context data, such as sequence length. In the latest GPT-3.5+ models, we find emergent abilities to generate seemingly random numbers and learn basic formal languages, with striking in-context learning dynamics where model outputs transition sharply from seemingly random behaviors to deterministic repetition.

As artificial intelligence models have exploded in scale and capability, understanding of their internal mechanisms remains a critical challenge. Inspired by the success of dynamical systems approaches in neuroscience, here we propose a novel framework for studying computations in deep learning systems. We focus on the residual stream (RS) in transformer models, conceptualizing it as a dynamical system evolving across layers. We find that activations of individual RS units exhibit strong continuity across layers, despite the RS being a non-privileged basis. Activations in the RS accelerate and grow denser over layers, while individual units trace unstable periodic orbits. In reduced-dimensional spaces, the RS follows a curved trajectory with attractor-like dynamics in the lower layers. These insights bridge dynamical systems theory and mechanistic interpretability, establishing a foundation for a "neuroscience of AI" that combines theoretical rigor with large-scale data analysis to advance our understanding of modern neural networks.

The fundamental units of internal representations in large language models (LLMs) remain undefined, limiting further understanding of their mechanisms. Neurons or features are often regarded as such units, yet neurons suffer from polysemy, while features face concerns of unreliable reconstruction and instability. To address this issue, we propose the Atoms Theory, which defines such units as atoms. We introduce the atomic inner product (AIP) to correct representation shifting, formally define atoms, and prove the conditions that atoms satisfy the Restricted Isometry Property (RIP), ensuring stable sparse representations over atom set and linking to compressed sensing. Under stronger conditions, we further establish the uniqueness and exact ell_1 recoverability of the sparse representations, and provide guarantees that single-layer sparse autoencoders (SAEs) with threshold activations can reliably identify the atoms. To validate the Atoms Theory, we train threshold-activated SAEs on Gemma2-2B, Gemma2-9B, and Llama3.1-8B, achieving 99.9% sparse reconstruction across layers on average, and more than 99.8% of atoms satisfy the uniqueness condition, compared to 0.5% for neurons and 68.2% for features, showing that atoms more faithfully capture intrinsic representations of LLMs. Scaling experiments further reveal the link between SAEs size and recovery capacity. Overall, this work systematically introduces and validates Atoms Theory of LLMs, providing a theoretical framework for understanding internal representations and a foundation for mechanistic interpretability. Code available at https://github.com/ChenhuiHu/towards_atoms.