Activation patching is a standard method in mechanistic interpretability for localizing the components of a model responsible for specific behaviors, but it is computationally expensive to apply at scale. Attribution patching offers a faster, gradient-based approximation, yet suffers from noise and reduced reliability in deep, highly non-linear networks. In this work, we introduce Relevance Patching (RelP), which replaces the local gradients in attribution patching with propagation coefficients derived from Layer-wise Relevance Propagation (LRP). LRP propagates the network's output backward through the layers, redistributing relevance to lower-level components according to local propagation rules that ensure properties such as relevance conservation or improved signal-to-noise ratio. Like attribution patching, RelP requires only two forward passes and one backward pass, maintaining computational efficiency while improving faithfulness. We validate RelP across a range of models and tasks, showing that it more accurately approximates activation patching than standard attribution patching, particularly when analyzing residual stream and MLP outputs in the Indirect Object Identification (IOI) task. For instance, for MLP outputs in GPT-2 Large, attribution patching achieves a Pearson correlation of 0.006, whereas RelP reaches 0.956, highlighting the improvement offered by RelP. Additionally, we compare the faithfulness of sparse feature circuits identified by RelP and Integrated Gradients (IG), showing that RelP achieves comparable faithfulness without the extra computational cost associated with IG.

We introduce combinatorial interpretability, a methodology for understanding neural computation by analyzing the combinatorial structures in the sign-based categorization of a network's weights and biases. We demonstrate its power through feature channel coding, a theory that explains how neural networks compute Boolean expressions and potentially underlies other categories of neural network computation. According to this theory, features are computed via feature channels: unique cross-neuron encodings shared among the inputs the feature operates on. Because different feature channels share neurons, the neurons are polysemantic and the channels interfere with one another, making the computation appear inscrutable. We show how to decipher these computations by analyzing a network's feature channel coding, offering complete mechanistic interpretations of several small neural networks that were trained with gradient descent. Crucially, this is achieved via static combinatorial analysis of the weight matrices, without examining activations or training new autoencoding networks. Feature channel coding reframes the superposition hypothesis, shifting the focus from neuron activation directionality in high-dimensional space to the combinatorial structure of codes. It also allows us for the first time to exactly quantify and explain the relationship between a network's parameter size and its computational capacity (i.e. the set of features it can compute with low error), a relationship that is implicitly at the core of many modern scaling laws. Though our initial studies of feature channel coding are restricted to Boolean functions, we believe they provide a rich, controlled, and informative research space, and that the path we propose for combinatorial interpretation of neural computation can provide a basis for understanding both artificial and biological neural circuits.

Autoregressive transformer language models (LMs) possess strong syntactic abilities, often successfully handling phenomena from agreement to NPI licensing. However, the features they use to incrementally process language inputs are not well understood. In this paper, we fill this gap by studying the mechanisms underlying garden path sentence processing in LMs. We ask: (1) Do LMs use syntactic features or shallow heuristics to perform incremental sentence processing? (2) Do LMs represent only one potential interpretation, or multiple? and (3) Do LMs reanalyze or repair their initial incorrect representations? To address these questions, we use sparse autoencoders to identify interpretable features that determine which continuation - and thus which reading - of a garden path sentence the LM prefers. We find that while many important features relate to syntactic structure, some reflect syntactically irrelevant heuristics. Moreover, while most active features correspond to one reading of the sentence, some features correspond to the other, suggesting that LMs assign weight to both possibilities simultaneously. Finally, LMs do not re-use features from garden path sentence processing to answer follow-up questions.

Recent advances in the understanding of neural networks suggest that superposition, the ability of a single neuron to represent multiple features simultaneously, is a key mechanism underlying the computational efficiency of large-scale networks. This paper explores the theoretical foundations of computing in superposition, focusing on explicit, provably correct algorithms and their efficiency. We present the first lower bounds showing that for a broad class of problems, including permutations and pairwise logical operations, a neural network computing in superposition requires at least $\Omega(m' \log m')$ parameters and $\Omega(\sqrt{m' \log m'})$ neurons, where $m'$ is the number of output features being computed. This implies that any ``lottery ticket'' sparse sub-network must have at least $\Omega(m' \log m')$ parameters no matter what the initial dense network size. Conversely, we show a nearly tight upper bound: logical operations like pairwise AND can be computed using $O(\sqrt{m'} \log m')$ neurons and $O(m' \log^2 m')$ parameters. There is thus an exponential gap between computing in superposition, the subject of this work, and representing features in superposition, which can require as little as $O(\log m'$) neurons based on the Johnson-Lindenstrauss Lemma. Our hope is that our results open a path for using complexity theoretic techniques in neural network interpretability research.

We introduce methods for discovering and applying sparse feature circuits. These are causally implicated subnetworks of human-interpretable features for explaining language model behaviors. Circuits identified in prior work consist of polysemantic and difficult-to-interpret units like attention heads or neurons, rendering them unsuitable for many downstream applications. In contrast, sparse feature circuits enable detailed understanding of unanticipated mechanisms. Because they are based on fine-grained units, sparse feature circuits are useful for downstream tasks: We introduce SHIFT, where we improve the generalization of a classifier by ablating features that a human judges to be task-irrelevant. Finally, we demonstrate an entirely unsupervised and scalable interpretability pipeline by discovering thousands of sparse feature circuits for automatically discovered model behaviors.