Deep neural networks are notorious for defying theoretical treatment. However, when the number of parameters in each layer tends to infinity, the network function is a Gaussian process (GP) and quantitatively predictive description is possible. Gaussian approximation allows one to formulate criteria for selecting hyperparameters, such as variances of weights and biases, as well as the learning rate. These criteria rely on the notion of criticality defined for deep neural networks. In this work we describe a new practical way to diagnose criticality.

Recent experiments reveal that task-relevant variables are often encoded in approximately orthogonal subspaces of the neural activity space. These disentangled low-dimensional representations are observed in multiple brain areas and across different species, and are typically the result of a process of abstraction that supports simple forms of out-of-distribution generalization. The mechanisms by which such geometries emerge remain poorly understood, and the mechanisms that have been investigated are typically unsupervised (e.g., based on variational auto-encoders). Here, we show mathematically that abstract representations of latent variables are guaranteed to appear in the last hidden layer of feedforward nonlinear networks when they are trained on tasks that depend directly on these latent variables. These abstract representations reflect the structure of the desired outputs or the semantics of the input stimuli. To investigate the neural representations that emerge in these networks, we develop an analytical framework that maps the optimization over the network weights into a mean-field problem over the distribution of neural preactivations. Applying this framework to a finite-width ReLU network, we find that its hidden layer exhibits an abstract representation at all global minima of the task objective. We further extend these analyses to two broad families of activation functions and deep feedforward architectures, demonstrating that abstract representations naturally arise in all these scenarios. Together, these results provide an explanation for the widely observed abstract representations in both the brain and artificial neural networks, as well as a mathematically tractable toolkit for understanding the emergence of different kinds of representations in task-optimized, feature-learning network models.

We establish that randomly initialized neural networks, with large width and a natural choice of hyperparameters, have nearly independent outputs exactly when their activation function is nonlinear with zero mean under the Gaussian measure: $\mathbb{E}_{z \sim \mathcal{N}(0,1)}[σ(z)]=0$. For example, this includes ReLU and GeLU with an additive shift, as well as tanh, but not ReLU or GeLU by themselves. Because of their nearly independent outputs, we propose neural networks with zero-mean activation functions as a promising candidate for the Alignment Research Center's computational no-coincidence conjecture -- a conjecture that aims to measure the limits of AI interpretability.

Deep neural networks are notorious for defying theoretical treatment.

Branch-and-bound with preactivation splitting has been shown highly effective for deterministic verification of neural networks. In this paper, we extend this framework to the probabilistic setting. We propose BaB-prob that iteratively divides the original problem into subproblems by splitting preactivations and leverages linear bounds computed by linear bound propagation to bound the probability for each subproblem. We prove soundness and completeness of BaB-prob for feedforward-ReLU neural networks. Furthermore, we introduce the notion of uncertainty level and design two efficient strategies for preactivation splitting, yielding BaB-prob-ordered and BaB+BaBSR-prob. We evaluate BaB-prob on untrained networks, MNIST and CIFAR-10 models, respectively, and VNN-COMP 2025 benchmarks. Across these settings, our approach consistently outperforms state-of-the-art approaches in medium- to high-dimensional input problems.

We propose a testable universality hypothesis, asserting that seemingly disparate neural network solutions observed in the simple task of modular addition are unified under a common abstract algorithm. While prior work interpreted variations in neuron-level representations as evidence for distinct algorithms, we demonstrate - through multi-level analyses spanning neurons, neuron clusters, and entire networks - that multilayer perceptrons and transformers universally implement the abstract algorithm we call the approximate Chinese Remainder Theorem. Crucially, we introduce approximate cosets and show that neurons activate exclusively on them. Furthermore, our theory works for deep neural networks (DNNs). It predicts that universally learned solutions in DNNs with trainable embeddings or more than one hidden layer require only O(log n) features, a result we empirically confirm. This work thus provides the first theory-backed interpretation of multilayer networks solving modular addition. It advances generalizable interpretability and opens a testable universality hypothesis for group multiplication beyond modular addition.

Mathematical reasoning is an increasingly important indicator of large language model (LLM) capabilities, yet we lack understanding of how LLMs process even simple mathematical tasks. To address this, we reverse engineer how three mid-sized LLMs compute addition. We first discover that numbers are represented in these LLMs as a generalized helix, which is strongly causally implicated for the tasks of addition and subtraction, and is also causally relevant for integer division, multiplication, and modular arithmetic. We then propose that LLMs compute addition by manipulating this generalized helix using the "Clock" algorithm: to solve $a+b$, the helices for $a$ and $b$ are manipulated to produce the $a+b$ answer helix which is then read out to model logits. We model influential MLP outputs, attention head outputs, and even individual neuron preactivations with these helices and verify our understanding with causal interventions. By demonstrating that LLMs represent numbers on a helix and manipulate this helix to perform addition, we present the first representation-level explanation of an LLM's mathematical capability.