Phase Transitions in Neural Networks
Latest papers
37 papers9 papers in the last four weeks, against 2 the four weeks before. 0.1% of all new papers.
Large language model compression has made substantial progress through pruning, quantization, and low-rank decomposition, yet a fundamental limitation persists across all existing methods: compression budgets are allocated without any representation of what individual model components functionally encode. We term this the capability-blind compression problem and argue it is a root cause of two well-documented failures -- the insensitivity of perplexity-based evaluation to reasoning capability loss, and the abrupt phase transitions in model performance recently characterized by Ma et al. (2026). We propose Capability-Guided Compression (CGC), a framework that addresses this by using Sparse Autoencoder (SAE)-derived capability density maps to allocate differential compression budgets across transformer components. Capability density is a formally defined scalar measure combining the feature breadth, activation entropy, and cross-input consistency of a component's SAE feature activation distribution. We prove theoretically that components with higher capability density exhibit lower structural redundancy and reach their individual phase transition points at lower compression ratios, providing the first pre-compression mechanism for component-level phase transition prediction. Experiments on GPT-2 Medium confirm that capability density is statistically independent of Wanda importance scores (Spearman rho = -0.054, n = 384 heads), establishing it as a genuinely novel compression signal orthogonal to all existing importance metrics. We report a negative result on PPL-based compression comparison and provide a principled diagnosis identifying GPT-2 Medium as an insufficient test bed for the full CGC hypothesis. The theoretical framework, density formalism, and orthogonality finding constitute a foundation for capability-aware compression research.
Pruning-induced phases in fully-connected neural networks: the eumentia, the dementia, and the amentia
Modern neural networks are heavily overparameterized, and pruning, which removes redundant neurons or connections, has emerged as a key approach to compressing them without sacrificing performance. However, while practical pruning methods are well developed, whether pruning induces sharp phase transitions in the neural networks and, if so, to what universality class they belong, remain open questions. To address this, we study fully-connected neural networks trained on MNIST, independently varying the dropout (i.e., removing neurons) rate at both the training and evaluation stages to map the phase diagram. We identify three distinct phases: eumentia (the network learns), dementia (the network has forgotten), and amentia (the network cannot learn), sharply distinguished by the power-law scaling of the cross-entropy loss with the training dataset size. {In the eumentia phase, the algebraic decay of the loss, as documented in the machine learning literature as neural scaling laws, is from the perspective of statistical mechanics the hallmark of quasi-long-range order.} We demonstrate that the transition between the eumentia and dementia phases is accompanied by scale invariance, with a diverging length scale that exhibits hallmarks of a Berezinskii-Kosterlitz-Thouless-like transition; the phase structure is robust across different network widths and depths. Our results establish that dropout-induced pruning provides a concrete setting in which neural network behavior can be understood through the lens of statistical mechanics.
Marginals Before Conditionals
We construct a minimal task that isolates conditional learning in neural networks: a surjective map with K-fold ambiguity, resolved by a selector token z, so H(A | B) = log K while H(A | B, z) = 0. The model learns the marginal P(A | B) first, producing a plateau at exactly log K, before acquiring the full conditional in a sharp, collective transition. The plateau has a clean decomposition: height = log K (set by ambiguity), duration = f(D) (set by dataset size D, not K). Gradient noise stabilizes the marginal solution: higher learning rates monotonically slow the transition (3.6* across a 7* η range at fixed throughput), and batch-size reduction delays escape, consistent with an entropic force opposing departure from the low-gradient marginal. Internally, a selector-routing head assembles during the plateau, leading the loss transition by ~50% of the waiting time. This is the Type 2 directional asymmetry of Papadopoulos et al. [2024], measured dynamically: we track the excess risk from log K to zero and characterize what stabilizes it, what triggers its collapse, and how long it takes.
Optimal scaling laws in learning hierarchical multi-index models
In this work, we provide a sharp theory of scaling laws for two-layer neural networks trained on a class of hierarchical multi-index targets, in a genuinely representation-limited regime. We derive exact information-theoretic scaling laws for subspace recovery and prediction error, revealing how the hierarchical features of the target are sequentially learned through a cascade of phase transitions. We further show that these optimal rates are achieved by a simple, target-agnostic spectral estimator, which can be interpreted as the small learning-rate limit of gradient descent on the first-layer weights. Once an adapted representation is identified, the readout can be learned statistically optimally, using an efficient procedure. As a consequence, we provide a unified and rigorous explanation of scaling laws, plateau phenomena, and spectral structure in shallow neural networks trained on such hierarchical targets.
Critical Organization of p-Adic Deep Neural Networks
We study continuous neural networks whose neurons are the points of the ring of p-adic integers, a fractal tree-like space that encodes hierarchical architectures. The hidden states solve a nonlinear integral equation, which describes the equilibria of recurrent networks and deep equilibrium models; feedforward architectures appear as the special case of kernels without feedback. We prove that the state is unique and depends Lipschitz-continuously on the parameters whenever the product of the Lipschitz constant of the activation function and the operator norm of the connectivity is less than one, and that, when the parameters are locally constant, every state is exactly the state of a finite network with a tree-like topology. We introduce p-adic discrete deep networks, whose layers are the levels of the tree and whose weights are coarse-grainings of a single kernel, and show that, in the stable region, they converge to the continuous network as the depth tends to infinity. Feedforward networks always have a unique state; hence a multiplicity of states, and therefore critical organization in our sense, requires recurrence. For recurrent networks with radial kernels we prove a sharp transition: the zero state is the only state if and only if the total coupling does not exceed one. For hierarchical kernels, above the threshold, the states are indexed by sign patterns on the leaves of the tree, their number is computed exactly, and they form a tree-indexed partially ordered set graded by instability. Numerical experiments illustrate these results and test conjectures on general radial kernels and on random recurrent networks on the tree. Finally, we interpret the critical organization hypothesis for deep neural networks, originally formulated for brain organization, through the lens of multistability, and we discuss the continuous limits of families of finite networks.
Toward a Physics of Deep Learning and Brains
Deep neural networks and brains both learn and share superficial similarities: processing nodes are likened to neurons and adjustable weights are likened to modifiable synapses. But can a unified theoretical framework be found to underlie them both? Here we show that the equations used to describe neuronal avalanches in living brains can also be applied to cascades of activity in deep neural networks. These equations are derived from non-equilibrium statistical physics and show that deep neural networks learn best when poised between absorbing and active phases. Because these networks are strongly driven by inputs, however, they do not operate at a true critical point but within a quasi-critical regime -- one that still approximately satisfies crackling noise scaling relations. By training networks with different initializations, we show that maximal susceptibility is a more reliable predictor of learning than proximity to the critical point itself. This provides a blueprint for engineering improved network performance. Finally, using finite-size scaling we identify distinct universality classes, including Barkhausen noise and directed percolation. This theoretical framework demonstrates that universal features are shared by both biological and artificial neural networks.
Storage and Learning phase transitions in the Random-Features Hopfield Model
The Hopfield model is a paradigmatic model of neural networks that has been analyzed for many decades in the statistical physics, neuroscience, and machine learning communities. Inspired by the manifold hypothesis in machine learning, we propose and investigate a generalization of the standard setting that we name Random-Features Hopfield Model. Here binary patterns of length are generated by applying to Gaussian vectors sampled in a latent space of dimension a random projection followed by a non-linearity. Using the replica method from statistical physics, we derive the phase diagram of the model in the limit with fixed ratios and . Besides the usual retrieval phase, where the patterns can be dynamically recovered from some initial corruption, we uncover a new phase where the features characterizing the projection can be recovered instead. We call this phenomena the learning phase transition, as the features are not explicitly given to the model but rather are inferred from the patterns in an unsupervised fashion.