Micah Adler

Micah Adler, Nir Shavit

Superposition, the ability of neural networks to represent more features than neurons, is increasingly seen as key to the efficiency of large models. This paper investigates the theoretical foundations of computing in superposition, establishing complexity bounds for explicit, provably correct algorithms. We present the first lower bounds for a neural network computing in superposition, showing that for a broad class of problems, including permutations and pairwise logical operations, computing $m'$ features in superposition requires at least $Ω(\sqrt{m' \log m'})$ neurons and $Ω(m' \log m')$ parameters. This implies an explicit limit on how much one can sparsify or distill a model while preserving its expressibility, and complements empirical scaling laws by implying the first subexponential bound on capacity: a network with $n$ neurons can compute at most $O(n^2 / \log n)$ features. Conversely, we provide a nearly tight constructive 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 the complexity of computing in superposition (the subject of this work) versus merely representing features, which can require as little as $O(\log m')$ neurons based on the Johnson-Lindenstrauss Lemma. Our work analytically establishes that the number of parameters is a good estimator of the number of features a neural network computes.

Shingo Kodama, Niv Cohen, Micah Adler et al.

While many recent methods aim to unlearn or remove knowledge from pretrained models, seemingly erased knowledge often persists and can be recovered in various ways. Because large foundation models are far from interpretable, understanding whether and how such knowledge persists remains a significant challenge. To address this, we turn to the recently developed framework of combinatorial interpretability. This framework, designed for two-layer neural networks, enables direct inspection of the knowledge encoded in the model weights. We reproduce baseline unlearning methods within the combinatorial interpretability setting and examine their behavior along two dimensions: (i) whether they truly remove knowledge of a target concept (the concept we wish to remove) or merely inhibit its expression while retaining the underlying information, and (ii) how easily the supposedly erased knowledge can be recovered through various fine-tuning operations. Our results shed light within a fully interpretable setting on how knowledge can persist despite unlearning and when it might resurface.

Micah Adler, Dan Alistarh, Nir Shavit

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.

Linghao Kong, Inimai Subramanian, Yonadav Shavit et al.

This work demonstrates how increasing the number of neurons in a network without increasing its number of non-zero parameters improves performance. We show that this gain corresponds with a decrease in interference between multiple features that would otherwise share the same neurons. To reduce such entanglement at a fixed non-zero parameter count, we introduce Fixed Parameter Expansion (FPE): replace a neuron with multiple children and partition the parent's weights disjointly across them, so that each child inherits a non-overlapping subset of connections. On symbolic tasks, specifically Boolean code problems, clause-aligned FPE systematically reduces polysemanticity metrics and yields higher task accuracy. Notably, random splits of neuron weights approximate these gains, indicating that reduced collisions, not precise assignment, are a primary driver. Consistent with the superposition hypothesis, the benefits of FPE grow with increasing interference: when polysemantic load is high, accuracy improvements are the largest. Transferring these insights to real models (classifiers over CLIP embeddings and deeper multilayer networks) we find that widening networks while maintaining a constant non-zero parameter count consistently increases accuracy. These results identify an interpretability-grounded mechanism to leverage width against superposition, improving performance without increasing the number of non-zero parameters. Such a direction is well matched to modern accelerators, where memory movement of non-zero parameters, rather than raw compute, is the dominant bottleneck.

Linghao Kong, Angelina Ning, Micah Adler et al.

A recently discovered class of entangled neurons, known as Wasserstein neurons, is disproportionately critical in large language models despite constituting only a very small fraction of the network: their targeted removal collapses the model, consistent with their unique role in differentiating similar inputs. Interestingly, in Wasserstein neurons immediately preceding smooth activation functions, such differentiation manifests in the negative pre-activation space, especially in early layers. Pairs of similar inputs are driven to highly distinct negative values, and these pairs involve syntactic tokens such as determiners and prepositions. We show that this negative region is functional rather than simply favorable for optimization. A minimal, sign-specific intervention that zeroes only the negative pre-activations of a small subset of entangled neurons significantly weakens overall model function and disrupts grammatical behavior, while both random and perplexity-matched controls leave grammatical performance largely unchanged. Part of speech analysis localizes the excess surprisal to syntactic scaffolding tokens, and layer-specific interventions reveal that small local degradations accumulate across depth. Over training checkpoints, the same ablation impairs grammatical behavior as Wasserstein neurons emerge and stabilize. Together, these results identify negative differentiation in a sparse subset of entangled neurons as a crucial mechanism that language models rely on for syntax.

Micah Adler

While self-attention is known to learn relations among tokens, we lack a formal understanding of its capacity: how many distinct relations can a single layer reliably recover for a given budget? To formalize this, we introduce Relational Graph Recognition (RGR), where the key-query channel represents a graph on $m$ items with $m'$ directed edges, and, given a context of items, must recover the neighbors of each item. We measure resources by the total key dimension $D_K = h\,d_k$. Within this framework, we analytically derive a capacity scaling law and validate it empirically. We show that $D_K = Θ(m' \log m' / d_{\text{model}})$ is both necessary (information-theoretic lower bound) and sufficient (explicit construction) in a broad class of graphs to recover $m'$ relations. This scaling law directly leads to a new, capacity-based rationale for multi-head attention that applies even when each item only attends to a single target. When embeddings are uncompressed ($m = d_{\text{model}}$) and the graph is a permutation, a single head suffices. However, compression ($m > d_{\text{model}}$) forces relations into overlapping subspaces, creating interference that a single large head cannot disentangle. Our analysis shows that allocating a fixed $D_K$ across many small heads mitigates this interference, increasing the number of recoverable relations. Controlled single-layer experiments mirror the theory, revealing a sharp performance threshold that matches the predicted capacity scaling and confirms the benefit of distributing $D_K$ across multiple heads. Altogether, these results provide a concrete scaling law for self-attention capacity and a principled design rule for allocating key-query budget across heads.

Shashata Sawmya, Micah Adler, Nir Shavit

This paper studies the emergence of interpretable categorical features within large language models (LLMs), analyzing their behavior across training checkpoints (time), transformer layers (space), and varying model sizes (scale). Using sparse autoencoders for mechanistic interpretability, we identify when and where specific semantic concepts emerge within neural activations. Results indicate clear temporal and scale-specific thresholds for feature emergence across multiple domains. Notably, spatial analysis reveals unexpected semantic reactivation, with early-layer features re-emerging at later layers, challenging standard assumptions about representational dynamics in transformer models.