Overview

Implicit bias is the tendency of a model and its training procedure to select some solutions or representations over others, even without an explicit regularizer; these preferences shape generalization, extrapolation, and sometimes fairness. Researchers derive these preferences from gradient descent, gradient flow, and stochastic optimization in settings such as separable classification, deep linear and ReLU networks, matrix factorization, and attention, often finding max-margin, sparse, or low-rank solutions. They vary initialization, depth, parameterization, optimizer, step size, and noise, and test consequences in tasks including length generalization, shortcut learning, and demographic representations in language and vision-language models. Current work is moving beyond asymptotic results in simplified models toward finite-time dynamics and realistic architectures, but the selected bias can shift during training and improve one notion of generalization while worsening another; fair outputs can also conceal biased internal representations.

Latest papers

36 papers7 papers in the last four weeks, up 75% on the four weeks before. 0.0% of all new papers.

Nov 12, 2026

Classification with Large-Step Gradient Descent: Unifying Self-Stabilization and Convergence Rates

Neural networks are often trained with gradient descent (GD) using step sizes that exceed classical stability thresholds, a regime where standard convergence theory no longer applies. Despite exhibiting non-monotonic loss and oscillatory dynamics, this regime frequently leads to faster convergence and better generalization. Existing theoretical perspectives offer complementary but incomplete insights: one identifies a self-stabilization mechanism that drives GD toward flat regions but does not yield convergence rates; another derives rigorous rates for specific losses but does not explain the underlying stabilization mechanism; and a third unifies these views for overparameterized least squares, but relies on the existence of a minimizer manifold, which is absent in classification settings. In this work, we extend this unified geometric perspective to exponential-tail losses, including logistic and cross-entropy losses in separable linear classification, where the loss infimum is not attained at any finite parameter. We identify a curvature-defined reference subspace that replaces the role of the minimizer manifold, yielding a coupled dynamical system in which orthogonal oscillations are damped by parallel progress along directions of decreasing sharpness. Under regularity conditions, which we verify for logistic and multiclass cross-entropy losses on symmetric data models, we derive explicit bounds for the iterates, not just the loss, covering both the unstable and stable regimes, and revealing a three-phase damping structure in the latter. Our analysis further uncovers a transient implicit bias induced by large step sizes: although GD converges asymptotically to the max-margin direction, large steps can leave persistent orthogonal residuals and, in certain geometries, amplify them, affecting finite-time alignment with the max-margin direction.
Nov 6, 2026

Implicit Bias in State Space Models and Linear Autoregressive Training

Linear State Space Models (SSMs) and linear autoregressive sequence models reuse the same parameters across a sample-dependent number of steps. As a result, different examples impose constraints through different predictors, even though all predictors share the same weights. We study this phenomenon through a general model of multiple non-homogeneous polynomial predictors with shared parameters, in a classification setting. Under standard interpolation and directional-convergence assumptions, together with a trajectory-dependent effective-degree condition, we show that gradient descent on the exponential loss converges in direction to a KKT point of an effective-degree max-margin problem. In this problem, only examples with minimal effective degree impose hard margin constraints; examples with faster-growing margins are asymptotically inactive except for feasibility. Our result provides a first-order characterization of the implicit bias in linear SSMs and linear autoregressive models. It exposes a mechanism of length bias: a longer input sequence or additional autoregressive steps may increase the example's effective degree, and the examples with the smallest effective degree dominate the limiting classifier. Our result highlights how non-homogeneity and parameter sharing alter classical homogeneous implicit-bias results.
Nov 5, 2026

Adversarial Training with Large Step Sizes: Implicit Bias, Optimization, and Sharpness

Adversarial Training (AT) is an important method for improving model robustness. We study the dynamics of deterministic Gradient Descent (GD) in AT. Most existing works in this direction focus on the small or vanishing step sizes regime, where GD resembles a continuous-time flow. However, such step sizes can make training inefficient and are rarely used in practice. In this paper, we study the AT dynamics of GD with large constant step sizes. We first study GD in adversarial linear logistic regression, focusing on two aspects: where the model parameters converge (implicit bias) and how quickly the losses decrease (optimization). We then turn to neural networks training and study how the sharpness of losses evolve during AT. Our results advance the understanding of the behaviors of GD with large step sizes in AT and highlight three key differences from its behavior in Standard Training (ST): (i) the implicit bias gap between AT and ST, previously established only for small step sizes, persists for arbitrarily large constant step sizes, (ii) different from ST, there is no uniform rate of convergence for the optimization error in AT, and (iii) sharpness can exhibit two-stage dynamics that differ from the Edge-of-Stability behavior observed in ST.
Nov 1, 2026

Watch Your Step: Adam's Margin Maximization Bias Depends on the Learning Rate Schedule

We study the implicit bias of Adam and its similarity to sign gradient descent on homogeneous models in binary classification tasks. Perhaps surprisingly, we uncover that these depend on the learning rate schedule. While previous works studied strongly decaying learning rate schedules η(t)=o(t1/L−1)η(t)=\mathrm o \left(t^{1/L-1}\right), where LL is the homogeneity degree, and showed ℓ∞\ell_\infty-margin-maximization biases, we focus on the more common complementary regime, including constant and weakly decaying schedules. For a polynomial schedule tαt^α and homogeneity degree LL, we show that when α>1L−12α> \frac 1L - \frac12, Adam has no trajectory-independent margin-maximization bias for deep linear networks (in the KKT sense): first, there is no non-trivial, locally Lipschitz, convex implicit regularizer in this setting, including any squared norm. Second, on width-11 networks, the result holds without requiring convexity, implying that any implicit regularizer for wide networks must be trivial when restricted to a width-11 network. Nevertheless, we positively characterize the implicit bias in this regime as maximizing a trajectory-dependent weighted-ℓ2\ell_2 margin for any LL-homogeneous model. The fundamental mechanism behind our results is convergence of Adam's preconditioner to a trajectory-dependent vector, which also allows showing that the limiting parameter direction depends on initialization. In the intermediate regime 1L−1≤α≤1L−12\frac 1L - 1 \leq α\leq \frac1L - \frac12 we prove an ℓ∞\ell_\infty bias, conditional on gradient directional convergence to a full-support vector. Of independent interest, we establish a loss decay rate for full batch Adam on general homogeneous models, whose leading order depends only on the second-moment parameter.
Oct 29, 2026

Implicit Bias of (Stochastic) Gradient Descent on Deep Linear Diagonal Networks

Modern machine learning models are often overparameterized, i.e., have more parameters than training data. This typically implies that the loss function admits multiple minimizers so that training methods (and their hyperparameter choices) have significant influence on the computed solution. We study this so-called implicit bias phenomenon in the context of training deep linear diagonal networks. Previous work has shown that the gradient flow for learning such networks converges to (almost) sparse solutions when initialized small. For an arbitrary number of layers, we extend this analysis to the practically more realistic (stochastic) gradient descent iterations. Our approach identifies (stochastic) gradient descent as a perturbed (stochastic) mirror gradient descent. This allows us to establish convergence and characterize the limiting solution, relating it to weighted ℓ1\ell_1-minimization. Moreover, we provide sufficient conditions on the step sizes and quantify the influence of the step sizes on the implicit bias.
Oct 26, 2026

RAGing Against Misogyny: Improving Implicit Misogyny Explanations

Implicit misogyny often requires background knowledge to be understood. To address this challenge, we propose a Retrieval-Augmented Generation (RAG) framework for detecting and explaining implicit misogyny in English and Italian, leveraging encyclopedic, slang and feminist knowledge sources. Using the ImplicIT-Mis and SBIC+ datasets, we evaluate the contribution of each knowledge source to detection and explanation quality, investigating also whether RAG addresses known LLM failure modes in implicit misogyny. Results show that RAG improves recall on ImplicIT-Mis, but does not improve detection performance on SBIC+. However, human evaluation shows that RAG-based explanations are preferred over no-RAG explanations in both languages. Feminist literature emerges as the most consistently useful source for detection and explanation.
Oct 26, 2026

Mapping the Solution-Set Structure of Grokking on Latin Squares

Grokking is a delayed transition from memorisation to generalisation. A network first fits observed examples while remaining near chance on held-out inputs, then generalises abruptly. Prior work has focused mainly on this delay, finding that the network's main loss becomes weight decay after the fitting phase. We name the solution to this optimization problem as the minimum-norm endpoint, and study which tasks make all minimum-norm endpoints generalise from partial observations. Specifically, we model pairwise input tasks as edge colourings of Kq,qK_{q,q}, with Latin colourings as our structured family, and classify each observed instance as all-good, mixed, or no-good by its endpoint set. Furthermore, we prove two contrasting results: (1) Finite-abelian-group families are asymptotically all-good from partial observations; (2) Under a uniform random Latin square, all-good tasks become asymptotically rare for observation-based learners, leaving mixed or no-good behaviour. Moreover, experiments show that relabellings of addition generalise despite their dense polynomial representations. Finally, on tasks of composite order, finite-training results suggest that predictors learn a coarse part of the output before the full label.
Oct 16, 2026

A Minimal ReLU Classifier on Non-Linearly Separable Data: Convergence and Implicit Bias at and beyond the Edge of Stability

Gradient descent (GD) with large learning rates can successfully optimize neural networks even though classical smoothness-based learning rate conditions are violated. Yet theory for this regime largely covers linear models or linearly separable data. We analyze a minimal setting outside both regimes. Specifically, a two-neuron ReLU network trained with logistic loss on 1D non-linearly separable data. The objective is nonconvex and nonsmooth, and GD exhibits edge-of-stability (EoS) behavior. We identify an active cone in parameter space that contains every perfect classifier, where the problem becomes piecewise logistic regression with a varying effective dataset. Within this cone, we prove a finite-time loss bound for constant learning rates and show that GD converges in direction to an explicit max-margin vector, while the orthogonal component converges to the unique minimizer of a non-strongly convex potential. We further establish catapult results where exponentially small loss is reached after a single loss spike. Furthermore, we propose an adaptive schedule that lets learning rates grow to infinity without leaving the active cone. We prove that this schedule achieves a loss rate of O(1/(Tlog⁡T)O(1/(T\log{T}), which is faster than the standard O(1/T)O(1/T) rate. Our theoretical results extend to local convergence on more general network and are validated on MNIST dataset. Overall, we show that this setup provides a tractable neural-network model for studying large and adaptive step sizes beyond linear predictors and separable data.
Oct 8, 2026

From Geometry to Generalization: Why Row Normalization Can Beat Adam and Muon

Different optimizers can fit the same training data while selecting classifiers with substantially different geometries, but whether this difference provably affects population performance remains unclear. We show that row-wise normalization can achieve strictly higher population accuracy than full-batch Adam, a proxy for random-reshuffling Adam, and exact-SVD Muon in high-dimensional multiclass classification. Under an isotropic Gaussian-cloud data model, this advantage arises because row normalization's class-wise Euclidean geometry asymptotically preserves the population decision-boundary directions, whereas Adam's coordinate-wise geometry and Muon's spectral geometry introduce nonvanishing distortions. Beyond isotropy, the advantage persists for full-batch training on class means with independently oriented class-mean and test-noise covariances. It holds for power-law spectra with class-mean exponent below one, even under heavily anisotropic test noise. When both covariances are diagonal and sufficiently close, the advantage over Adam can reverse, while applying the same random rotation to both restores it by changing only their alignment with Adam's coordinate axes. Synthetic and last-layer language-model experiments support the predicted advantage.
Oct 5, 2026

The Implicit Bias of Hyperbolic Representation Learning for Multiclass Data: A Busemann Risk Perspective

We study the implicit bias of Riemannian gradient flow for hyperbolic multiclass classification with fixed class prototypes in hyperbolic space Hn\mathbb H^n. Our framework accommodates general permutation invariant relative margin (PERM) losses, a class that includes cross entropy and other standard multiclass losses. Our analysis is based on a decomposition: at large radius, the distance to each prototype splits into a radial term and a direction dependent term described by the Busemann function. This yields two main results. First, we prove a radial dichotomy: the sign of a drift coefficient μμ determines whether the radius is pushed toward the ideal boundary or back toward the interior; if the positive drift persists, then r(t)=12log⁡t+O(1)r(t)=\tfrac12\log t+O(1), while persistent negative drift returns the trajectory to the large radius threshold in finite time. Second, we show that large-radius directions can approach critical points of the Busemann risk on ∂Hn\partial\mathbb H^n, even during radial return. These results provide a rigorous large-radius perspective on two phenomena we refer to as boundary saturation and near-boundary clustering in hyperbolic representation learning.
Sep 30, 2026

Learning the identity: a case study of how SGD selects among functional decompositions

One might think that learning the identity function with a deep linear residual network is trivial - the path along residual connections already implements the identity, and so the network need only drive its weights to zero. However, this zero-weight solution is just one point on an entire manifold of population-loss minimizers, each corresponding to a different decomposition of the identity across the network's layers. Although the population loss does not distinguish among these solutions, stochastic gradient descent (SGD) reproducibly favors particular ones. For instance, under anisotropic label noise, the learned layers exhibit a noise-dependent spectrum; even with weight decay, SGD does not generally recover the zero-weight solution. Changing only the parametrization, while leaving the set of realizable functions unchanged, yields different behavior: factoring each weight matrix as a product of two matrices causes the weights to collapse to zero, even without explicit weight decay. While perhaps mysterious and unintuitive at first, these phenomena can be understood through the lens of entropic loss, which augments the population loss with a term proportional to the expected squared norm of the minibatch gradient (Ziyin et al., 2025). On the identity manifold, the population loss is constant, while the entropic term distinguishes among these decompositions. We characterize its minimizers analytically and use them to derive predictions for the structure of solutions favored by SGD. Networks trained with SGD closely match these predictions. Overall, the identity learning task studied here serves as a clean and simple case study of how the lens of entropic loss can clarify why SGD favors particular decompositions of the same input-output function.
Sep 27, 2026

On the Two Faces of Adam in Separable Linear Classification

We consider the behavior of deterministic, full-batch, bias-corrected Adam in separable linear classification with softmax parametrization under log-loss. In this setting, under a wide range of conditions Adam is known to approach max-norm-margin optimality when its stability constant εε is zero, while with a positive εε, it is known to approach Euclidean-margin optimality. Our main contribution is the quantitative description of Adam's behavior for small fixed positive εε. We give sufficient conditions under which an Adam-trained classifier nearly maximizes the max-norm margin before the updates become gradient-like. We also show that the classifier reaches a fixed target Euclidean margin only much later. Specifically, we show that for polynomially decreasing stepsizes with exponent aa, where 1/3<a<11/3<a<1, the updates become approximately proportional to the negative gradient after Θ(log⁡(1/ε)1/(1−a))Θ(\log(1/ε)^{1/(1-a)}) iterations. At that time, the classifier still nearly maximizes the max-norm margin. Reaching a fixed target Euclidean margin above that of every max-norm-optimal classifier, but below the optimum, is shown to require ε−Θ(1)/(1−a)ε^{-Θ(1)/(1-a)} iterations. Under inverse-linear stepsize decay (a=1a=1), the update transition takes polynomially many iterations, whereas reaching the target margin takes exponentially many. Experiments support these predictions. The later change in the classifier can improve or worsen generalization after training error reaches zero, connecting the analysis to grokking and its reverse.
Sep 3, 2026

High-Dimensional Learning Dynamics of Attention-Indexed Models

Attention mechanisms are central to modern foundation models, yet their training dynamics remain poorly understood, especially when the attention matrices have extensive rank. In this work, we study attention-indexed models, a broad framework that can represent multi-layer and multi-head attention architectures. First, we show that, in a suitable high-dimensional limit, the population-loss landscape is characterized by a finite set of trace order parameters. In contrast, online stochastic gradient descent (SGD) is governed by an infinite hierarchy of matrix moments, which we show can be exponentially well-approximated by a finite truncated system. Second, this framework reveals that attention parameterization itself can act as an architectural implicit bias. Direct optimization of an attention matrix S∈Rd×dS\in\mathbb{R}^{d\times d} can remain trapped in an uninformative state. Tied attention (S=WW⊤S=WW^\top) induces an automatic symmetry-breaking mechanism and yields weak recovery in Θ(d2log⁡d)Θ(d^2\log d) samples. For untied attention, S=UV⊤S=UV^\top, we uncover a fast-slow mechanism: the pre-activation mean first evolves on a fast timescale, while the overlaps evolve on a slower one. Weak recovery on the Θ(d2log⁡d)Θ(d^2\log d) scale occurs when the state selected by the fast dynamics breaks the initial symmetry.
Aug 5, 2026

Non-asymptotic implicit bias of logistic regression at early-stage gradient descent dynamics

Gradient descent has been of particular interest in modern machine learning beyond sole focus on optimization. Implicit bias emerging from optimization, though not being encoded by the learning objective, often prevents from overfitting to spurious patterns. A typical instance is the max-margin implicit bias of a linear classifier, widely established for exponentially tailed loss functions. Even after having a given dataset separated, the parameter vector continues to evolve towards the max-margin direction asymptotically along the gradient descent dynamics. This phenomenon corroborates a frequent empirical observation of "train longer, generalize better." However, the max-margin convergence is an asymptotic phenomenon, and what is worse, this asymptotic convergence rate is significantly slower than pure convex optimization. Even so, the parameter vector along gradient descent dynamics commonly correlates with the max-margin direction positively (though not exactly) within considerably fewer iterations than the asymptotic rate. By shedding another light on this classical problem, this work aims to understand the mechanism of this early-stage alignment phenomenon. Our theoretical results demonstrate that the parameter vector weakly aligns with the max-margin direction within O(exp⁡(exp⁡(−δ)))O(\exp(\exp(-δ))) iterations, where δ>0δ>0 is the permissible alignment error, which is shown to be tight. By tracking the radial and tangential flows, our proof operates on the alignment dynamics directly with dataset geometry and gets rid of the asymptotic expansion, which is a key insight to establishing faster weak alignment.
Aug 4, 2026

Variable Selection in the Context of AI Fairness

Fairness in AI systems has become more important with recent regulatory demands, such as the EU AI Act. Traditional approaches often do not take into account philosophical ethics and social awareness. Variable selection processes, in particular, can introduce implicit bias, affecting equity across different subgroups. We discuss a mathematical approach that evaluates fairness in AI, aligning mathematical methodologies with ethical considerations and regulatory requirements. Our aim is to advocate for interdisciplinary collaboration to address fairness, emphasizing the importance of understanding broader ethical and societal contexts. Our approach emphasizes maintaining all potentially relevant variables to allow for more granular fairness assessments and to reduce implicit bias. The findings suggest that the exclusion of sensitive or critical variables may compromise equity between subgroups. In contrast, retaining all relevant variables could reduce implicit bias. Thus, the interdisciplinary approach could provide deeper insight into the ethical implications and compliance with regulatory standards. By integrating a mathematical approach with ethical and social awareness, we suggest more equitable outcomes and responsible AI deployment. This work underscores the necessity of interdisciplinary collaboration in effectively addressing fairness in AI systems aligned with the objectives of the European Union's AI Act, which seeks to promote trustworthy and fair AI systems.
Jul 30, 2026

From Minds to Models: The Intersection of Psychology and LLM Behaviours

Large language models (LLMs) are often compared with the human mind because their decision-making is complex, non-linear and difficult to interpret. Psychological methods developed to investigate unobservable mental processes may therefore help examine LLM behaviour, particularly in government and healthcare. Building on prompt-based adaptations of the Implicit Association Test, this study tested whether ChatGPT produced sentiment differences across racial conditions in open-ended text. Fourteen base questions were crossed with eight racial categories and a race-agnostic control, producing 126 prompts. Each was submitted once to GPT-3.5T, GPT-4 and GPT-4T, yielding 378 responses. Sentiment scores were derived from categorical labels and source scores: positive labels retained the source score, negative labels were assigned its negative, and neutral responses were coded zero. A two-way ANOVA found a small main effect of racial condition, F(8, 351) = 2.04, p = .042, partial-eta squared = .044, but no effect of model, F(2, 351) = 0.07, p = .933, and no interaction, F(16, 351) = 0.23, p = .999. However, the effect was not retained in a rank-transformed sensitivity analysis, F(8, 351) = 1.53, p = .145, and Tukey-corrected comparisons found no significant pairwise differences. An uncorrected European-Indigenous Australian comparison was significant, but was selected post hoc and is reported only as hypothesis-generating. Evidence for sentiment differences was therefore weak and analysis-dependent. Sentiment scoring also cannot distinguish evaluative bias from the valence of historical content elicited by a prompt. We outline design changes needed to address these limitations and argue for interdisciplinary development of behavioural measures of model bias. Keywords: Implicit Bias, Psychological Research Methods, Artificial Intelligence, ChatGPT, Large Language Models, Sentiment Analysis
Jul 21, 2026

Relative Positions Generalize, Absolute Positions Memorize: An Implicit-Bias Account of Length Generalization in Attention

Transformers with relative positional encodings often extrapolate to sequences longer than those seen during training, whereas transformers with learned absolute encodings typically do not. This is a robust empirical regularity, and the explanations offered for it so far are chiefly about expressivity, that is, about whether a length-generalizing solution exists. We give an optimization explanation. On a minimal fixed-offset retrieval task that isolates positional selection, the gap is governed by the implicit bias of the trained attention head: among the many solutions that fit short sequences, which one gradient descent actually selects. We prove that rotary encodings make the attention logit a function of relative offset alone, an exact equivariance, so whatever selection rule is learned at training lengths is reproduced verbatim at every longer length. Learned absolute encodings instead leave out-of-range positions unconstrained, and the trained head pins to a fixed absolute position inside the training range. We characterize the learned rotary rule as a low-rank ``carrier'' kernel aligned with the target offset, and we derive the resulting graceful accuracy decay as an attention-dilution law; both predictions are confirmed across seeds and offsets. A linear-attention control shows the mechanism is specific to softmax: without normalization, training selects a min-norm interpolant that does not extrapolate. The phenomenon, the equivariance, and the carrier all transfer to a multi-layer, multi-head transformer trained on a full-sequence length-generalization task. The account connects the implicit bias of attention, implicit bias for extrapolation in recurrent models, and the learning side of the RASP-L conjecture.
Jul 14, 2026

Gradient Flow Dynamics and Implicit Bias of Diagonal Linear Networks under Infinitesimal Initialization

We study the gradient flow dynamics of diagonal linear networks for regression tasks under infinitesimal initialization. Extending Theorem 1 from Pesme & Flammarion (2023), we generalize the analysis to both deep diagonal linear networks and a broader class of two-layer diagonal linear networks (as defined in Definition 4.1). Specifically, we demonstrate that the training trajectories of these models can be equivalently characterized by the proposed Algorithm 1. We further prove that this algorithm converges to the solution of a modified l1\mathcal{l}_1 norm minimization problem. As a result, we establish that the implicit bias of both network architectures corresponds to a modified l1\mathcal{l}_1 norm in the regime of infinitesimal initialization. Additionally, we provide insights into the underlying mechanisms governing these dynamics by identifying the Structural Invariant Manifold (SIM) (Zhao et al., 2026) as the key geometric structure that shapes the learning process.
Jul 8, 2026

The Anatomy of Implicit Bias: Information Allocation in Neural Network Training

Implicit bias is usually explained as the preference of an optimization process for certain final solutions and their geometry. This view helps explain where a model finally stops. It gives less direct explanation of how this bias is formed during training. This paper proposes a training-time information allocation view. Under this view, optimization forms a writing pattern for error signals across parameter paths, coordinate channels, and sample regions. This paper builds a set of observable allocation diagnostics. These diagnostics include gradient demand, actual update injection, coordinate gain induced by exponential moving averages, channel-level update ratios, and sample-wise loss distributions. To separate training progress from internal allocation, this paper introduces a collapse--persistence analysis. Under matched training loss, if external loss statistics collapse but internal allocation ratios remain separated, then the factor changes the internal allocation of the training signal. Overall, this paper extends the analysis of implicit bias from final-solution geometry to training-time signal allocation. The main claim is that implicit bias is not only reflected by the final solution. It is also reflected by which parameter paths, coordinate channels, and sample regions receive the error signal first and more strongly during training. Based on this view, this paper places different training factors into a unified information-allocation diagnostic framework. The framework gives a mechanism-level explanation of training-time implicit bias. It also provides a basis for future optimization methods that control training progress and signal allocation separately.
Jul 6, 2026

The Map Behind the Flow: Finite-Step Gradient Descent as a Dynamical System

Many phenomena of deep learning are dynamical: they concern not only which minima exist, but how gradient descent reaches, avoids, or selects among them. Edge-of-stability behavior, sharpness oscillations, catapult phases, balancing, and movement toward flatter representations are effects of the training map itself, and are poorly captured by the small-step gradient-flow limit. This paper studies fixed-step gradient descent as a discrete dynamical system in a hierarchy of exactly solvable models retaining basic structures of deep learning: depth, factorization, width, data coupling, activation, and stochasticity. The starting point is the balanced scalar reduction of a deep linear chain, giving a quartic loss and a cubic gradient map whose post-edge behavior is explicit. Under the natural large-depth scaling, this dynamics converges to a universal Ricker-type map. The edge of stability is therefore not a breakdown of optimization, but the first bifurcation of the training map. Embedding the scalar dynamics back into factored models turns these regimes into learning phenomena. Finite steps break conservation laws of gradient flow and contract factorization imbalance; residual oscillations move parameters toward flatter, more balanced representations. Wider linear networks produce a ladder of spectral edges, so the optimal learning rate can lie beyond the first edge. Data coupling, nonlinear activations, and stochastic targets preserve the same organizing principle: finite-step oscillations drive alignment, balancing, and representation selection. Thus the learning rate is not merely a numerical stability parameter. It is a structural parameter of the training dynamics, determining its attractors and shaping the representations gradient descent selects.
Jul 3, 2026

Implicit Bias of SGD in Multivariate ReLU Networks: Effective Width Collapse

We study the implicit bias of noisy stochastic gradient descent in training wide two-layer ReLU networks for multivariate regression. In a mean-field regime, the training dynamics are approximated by a Wasserstein gradient flow that converges to a unique stationary measure. We characterize the structure of this stationary measure and the predictor it represents. We show that, despite the network being infinitely overparameterized, the learned predictor admits an effectively finite representation: the input weights and biases align along finitely many directions, leading to an effective width collapse. In particular, the solution function is continuous piecewise affine, with affine regions determined by the cells of a finite hyperplane arrangement. The number of learned directions, and hence hyperplanes, is bounded above by 2P−12\mathcal{P}-1, where P\mathcal{P} denotes the number of linear dichotomies realizable on the training inputs. We further establish a non-redundancy property of the learned representation by proving that each learned direction induces a unique ternary activation pattern on the training data. Consequently, the complexity of the learned predictor is governed by the combinatorial geometry of the training data.
Jun 18, 2026

Fisher-Geometric Sharpness and the Implicit Bias of SGD toward Flat Minima

A widely held intuition in deep learning is that stochastic gradient descent (SGD) implicitly favors flat minima and that flat minima generalize better, but standard Euclidean measures of flatness such as the trace or maximum eigenvalue of the loss Hessian are not invariant under reparametrizations that preserve the network function, which undermines the theoretical foundations of this narrative. In this study we resolve this issue by grounding flatness in the Riemannian geometry of the statistical manifold induced by the Fisher Information Matrix (FIM). We define Riemannian sharpness mathematically and prove that it is invariant under smooth, function-preserving reparametrizations, which directly addresses the critique of Dinh et al. in the paper ``Sharp minima can generalize for deep nets''.We note that this invariance is a property of the true FIM; the diagonal empirical estimator used in practice (and in all experiments below) inherits invariance only approximately, and exact invariance under arbitrary reparametrizations would require structured estimators such as K-FAC. We formalize the gradient noise of mini-batch SGD as having a covariance structure proportional to the FIM, derive the stationary distribution of the resulting stochastic differential equation, and then show that the probability mass is exponentially concentrated at Riemannian-flat minima. A PAC-Bayes generalization bound controlled explicitly by SR formally links this geometric bias to test performance. Our experiments on MNIST and CIFAR-10 confirm that SR reliably tracks generalization in ways that Euclidean sharpness does not, and that its scaling with η/Bη/B matches the theoretical predictions. Together these results provide a rigorous, reparametrization-invariant account of why flat minima generalize.
May 29, 2026

Vision-Language Models Suppress Female Representations Under Ambiguous Input

Alignment teaches vision-language models (VLMs) to avoid expressing demographic biases, and when gender is clearly visible they largely succeed. Far less is known about ambiguous inputs (a worker in full gear, a figure seen from behind), cases common in practice yet rarely studied. We find that minimal prompting pressure exposes occupation-gender defaults when prompting ambiguous input images, with models collapsing to male even for strongly female-stereotyped occupations. But do these outputs reflect what models actually encode internally? We introduce LALS (Latent Association Leaning Score), a zero-shot metric that projects visual-token activations into the model's text-embedding space to measure concept associations per token and layer. Across 15 occupations, over 800 gender-ambiguous images, and four VLMs, internal representations and outputs often become systematically decoupled: models often encode a female association internally yet output male. Layer-wise analysis reveals an asymmetric filter: male signal amplifies end-to-end while female signal peaks mid-network and is suppressed before generation. A color ablation shows that culturally loaded visual cues such as clothing color further modulate these internal associations.
May 21, 2026

The Implicit Bias of Depth: From Neural Collapse to Softmax Codes

Neural collapse (NC) describes the structured geometry that emerges in the features and weights of trained classifiers. Recent theory suggests NC can be suboptimal in deep architectures, attributing this to an explicit low-rank bias from L2 regularization. We study the deep unconstrained feature model (UFM)-equivalent to a deep linear network with orthogonal inputs-trained without regularization, to isolate how gradient descent and depth alone shape NC. We show that depth induces an implicit low-rank bias: low-rank matrices propagate norm more efficiently through successive multiplications, promoting low-rank alternatives to NC. These alternatives, we argue, correspond to softmax codes: max-margin solutions previously found in width-bottlenecked networks. Analyzing training dynamics under spectral initialization, we identify an early-time repulsion among singular values that drives low-rank emergence, and characterize how depth shrinks NC's basin of attraction. Finally, we show that some effects act in the opposite direction: for randomly initialized networks, increasing width biases training toward higher-rank solutions. Our results provide the first asymptotic and dynamic characterization of implicit bias in deep UFMs trained with unregularized multiclass cross-entropy.
May 19, 2026

Implicit Bias of Mirror Flow in Homogeneous Neural Networks: Sparse and Dense Feature Learning

We study the max-margin solutions reached by mirror flow in deep neural networks with homogeneous activation functions. Extending classical results on gradient flow, we derive a novel balance equation for mirror flow from convex duality, enabling a characterization of the horizon function governing the induced margin. We further establish max-margin characterizations together with convergence rates and norm growth estimates. Finally, we support our theory through experiments on synthetic datasets and standard vision tasks. Concretely, we show that: (1) distinct non-homogeneous mirror maps can induce the same max-margin solution; (2) convergence can be extremely slow, including exponentially slow regimes; and (3) although all considered mirror maps exhibit feature learning, they can produce markedly different representations, ranging from sparse to dense neuron activations. Together, these results provide a unified perspective on sparse and dense feature learning in homogeneous neural networks, highlighting how mirror maps shape both optimization dynamics and the geometry of the learned classifiers.
May 12, 2026

Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions

Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconnect: models show no output-level bias, yet retain and amplify demographic representations across model layers. Through activation steering and novel cross-layer interventions, we demonstrate that this suppressed information is decision-relevant: when reinjected at critical layers, it produces near-complete decision reversals. Critically, this latent bias is asymmetric - steering interventions affect decisions in one demographic direction, while producing minimal effects in reverse - and susceptible to adversarial prompt engineering and parameter-efficient fine-tuning. These findings demonstrate that behavioural audits focused on outputs are insufficient: fair outputs can mask exploitable internal biases. They also motivate dual-layer testing frameworks combining output evaluation with representational analysis for AI governance in high-stakes decisions.
May 6, 2026

Estimating Implicit Regularization in Deep Learning

Deep learning systems are known to exhibit implicit regularization (alt. implicit bias), favoring simple solutions instead of merely minimizing the loss function. In some cases, we can analytically derive the implicit regularization -- connecting it to an equivalent penalty that augments the learning objective. However, modern deep learning systems are complex, carrying modifications to the training procedure and architecture (e.g. early stopping, minibatching, dropout) whose effects are not always directly interpretable. Although estimating the resulting implicit regularization could aid theorists in algorithm design and practitioners in interpreting their hyperparameter choices, this problem has received little direct attention. It is also tractable: regularization makes weight updates deviate from loss gradients, promising a signal for identifying implicit bias. Here we provide gradient matching methods that can be used to empirically estimate the implicit regularization. Our method works on networks with known regularization, recovering popular explicit penalties like ℓ1\ell_1 and ℓ2\ell_2. It also replicates known implicit effects, like the quadratic weight penalty induced by early stopping in gradient descent, demonstrating that it can be used to test theories of implicit regularization. Crucially, because our method is empirical, it can handle implicit regularization in arbitrary networks. We demonstrate this use by characterizing implicit ℓ2\ell_2 effects of dropout in deep networks. Our work shows that practitioners can use gradient matching to understand regularization in networks with implicit biases that are too complicated to derive analytically.
Mar 24, 2026

Towards The Implicit Bias on Multiclass Separable Data Under Norm Constraints

Implicit bias induced by gradient-based algorithms is essential to the generalization of overparameterized models, yet its mechanisms can be subtle. This work leverages the Normalized Steepest Descent} (NSD) framework to investigate how optimization geometry shapes solutions on multiclass separable data. We introduce NucGD, a geometry-aware optimizer designed to enforce low rank structures through nuclear norm constraints. Beyond the algorithm itself, we connect NucGD with emerging low-rank projection methods, providing a unified perspective. To enable scalable training, we derive an efficient SVD-free update rule via asynchronous power iteration. Furthermore, we empirically dissect the impact of stochastic optimization dynamics, characterizing how varying levels of gradient noise induced by mini-batch sampling and momentum modulate the convergence toward the expected maximum margin solutions.Our code is accessible at: https://github.com/Tsokarsic/observing-the-implicit-bias-on-multiclass-seperable-data.
Mar 11, 2026

Dual Space Preconditioning for Gradient Descent in the Overparameterized Regime

In this work, we study the convergence properties of the Dual Space Preconditioned Gradient Descent, encompassing optimizers such as Normalized Gradient Descent and Gradient Clipping. We consider preconditioners of the form ∇K\nabla K, where K:Rd×k→RK: \mathbb{R}^{d \times k} \to \mathbb{R} is convex and apply ∇K(⋅)\nabla K(\cdot) to train an over-parameterized linear model with a convex loss of the form ℓ(XW−Y)\ell(X W - Y), for weights W∈Rd×kW \in \mathbb{R}^{d \times k}, labels Y∈Rn×kY \in \mathbb{R}^{n \times k} and data X∈Rn×dX \in \mathbb{R}^{n \times d}. Under the aforementioned assumptions, we prove that the iterates of the full-batch preconditioned gradient descent converge at an exponential rate to a point W∞∈Rd×kW_{\infty} \in \mathbb{R}^{d \times k} satisfying XW∞=YXW_{\infty} = Y. We also study the implicit bias of Dual Space Preconditioned Gradient Descent. First, we demonstrate analytically and empirically that, for general K(⋅)K(\cdot), W∞W_\infty depends on the chosen constant step size, hindering a precise characterization of the implicit bias. We also provide an approximate implicit bias property for general preconditioners, namely, ∥W0−W∞∥F≤c∥W0−WGD,∞∥F\|W_0 - W_{\infty}\|_F \le c \|W_0 - W_{\text{GD}, \infty}\|_F for a constant c>0c>0 and WGD,∞W_{\text{GD}, \infty} denoting the convergence point of GD initialized at W0W_0. Furthermore, for preconditioners of the form K(G)=h(∥G∥F)K(G) = h(\|G\|_F), known as {\it isotropic preconditioners}, and for the stochastic variation of the algorithm with arbitrary batch-size, we prove linear convergence to WGD,∞W_{\text{GD}, \infty}. Finally, in the experiments, we demonstrate faster convergence on a nonlinear model obtained using the smoothed matrix elastic-net as a preconditioner.
Mar 9, 2026

Minor First, Major Last: A Depth-Induced Implicit Bias of Sharpness-Aware Minimization

We study the implicit bias of Sharpness-Aware Minimization (SAM) when training LL-layer linear diagonal networks on linearly separable binary classification. For linear models (L=1L=1), both ℓ∞\ell_\infty- and ℓ2\ell_2-SAM recover the ℓ2\ell_2 max-margin classifier, matching gradient descent (GD). However, for depth L=2L = 2, the behavior changes drastically -- even on a single-example dataset. For ℓ∞\ell_\infty-SAM, the limit direction depends critically on initialization and can converge to 0\mathbf{0} or to any standard basis vector, in stark contrast to GD, whose limit aligns with the basis vector of the dominant data coordinate. For ℓ2\ell_2-SAM, we show that although its limit direction matches the ℓ1\ell_1 max-margin solution as in the case of GD, its finite-time dynamics exhibit a phenomenon we call "sequential feature amplification", in which the predictor initially relies on minor coordinates and gradually shifts to larger ones as training proceeds or initialization increases. Our theoretical analysis attributes this phenomenon to ℓ2\ell_2-SAM's gradient normalization factor applied in its perturbation, which amplifies minor coordinates early and allows major ones to dominate later, giving a concrete example where infinite-time implicit-bias analyses are insufficient. Synthetic and real-data experiments corroborate our findings.