Simplicity Bias: Concepts and Implications
- Simplicity bias encompasses a variety of representations, ranging from low predictive-memory requirements to few decision-boundary pieces and minimal parameter counts, all of which impact a model's learning accuracy and robustness.
- Quantum and classical representations of simplicity can differ, leading to varied decision-making frameworks. Quantum representations, for example, show a different minimum spectral entropy
- A plethora of methods exist to alleviate the adverse effects of simplicity bias, which can include equalizing Predictive Power or diversifying the inclined sets of models, even though some of these techniques are context-focused remedies.
- Outcomes tend to be hugely dependent on the choice of representation, metric, optimization procedure, and evaluation context.
- A central open problem is developing methods that correctly identify when the model is prioritizing simplicity, which may represent robustness-beneficial feature clusters, with less reliance on less-probable parts of the dataset.
Simplicity bias denotes a family of inductive preferences through which learning systems favor comparatively simple representations, features, functions, or statistical explanations over alternatives that may be more complex, robust, equitable, or semantically appropriate. The term has no single framework-independent definition. Depending on the research program, simplicity may mean low predictive-memory requirements, few decision-boundary pieces, low-dimensional input dependence, low effective rank, low token sensitivity, short algorithmic description, low-order statistical interactions, sparse logical structure, or low variation along data-supported paths. Across these formulations, simplicity bias is not equivalent to correctness, robustness, fairness, causality, or generalization. Its effects depend on the representational framework, architecture, optimization procedure, data distribution, and evaluation objective.
1. Conceptual foundations and framework dependence
Simplicity is representation-relative. “The Ambiguity of Simplicity” distinguishes a framework-relative notion from an absolute notion such as minimal internal structure or memory. For stationary, ergodic stochastic processes, computational mechanics defines an -machine as the minimal unifilar predictor. Its causal states are equivalence classes of pasts that induce identical future predictions, and its classical statistical complexity is
The excess entropy,
measures predictive information shared between past and future, whereas measures the memory required by a minimal predictive model. They satisfy and coincide only in special cases (Aghamohammadi et al., 2016).
A quantum representation can encode distinct classical causal states as nonorthogonal quantum states. For stationary probabilities , the quantum memory state is
with quantum statistical complexity
The bounds
show that quantum memory can be strictly smaller than classical predictive memory. More importantly, the ordering between two processes can reverse: while 0. For the one-dimensional Ising chain, classical complexity is monotonic in temperature, while quantum complexity is generally smaller and nonmonotonic, reaching a maximum near 1 in the reported parameter setting. Consequently, simplicity can define only a partial ordering rather than a universal total ordering (Aghamohammadi et al., 2016).
This framework dependence extends to model selection. A complexity penalty based on 2 can prefer a different model from one based on 3. The same issue applies to Akaike-type information criteria, Bayesian Information Criterion, Minimum Description Length, Minimum Message Length, and Bayesian priors. If
4
then a simplicity-sensitive prior can change when the representational primitives change. The relevant subjectivity is not arbitrary personal taste: 5, 6, and 7 are mathematically defined, but the choice of framework and resource measure is not generally forced by the physical process alone.
Algorithmic-information approaches use Kolmogorov complexity,
8
and relate it to algorithmic probability through
9
Because exact Kolmogorov complexity is uncomputable, empirical work uses compression-based upper bounds. In discretized World Bank time series, a Lempel–Ziv-derived estimate 0 was used to test
1
For binary shape patterns, relative complexity predicted which of two patterns was more probable with reported success rates of 2–3 for high/low and up/down symbolizations, compared with a 4 baseline. This is a pattern-probability result, not numerical forecasting of future values (Dingle et al., 2022).
A broader Kolmogorov-complexity interpretation connects simplicity bias to the No Free Lunch theorem. Uniformly sampled labelings are almost always incompressible, so no bounded-complexity learner can generalize on them. Real-world datasets are atypically structured and compressible. Experiments reported that CNNs could compress reshaped tabular datasets lacking meaningful spatial structure, while pre-trained and randomly initialized LLMs preferentially generated low-complexity sequences. This supports a distributional interpretation in which neural networks and naturally occurring data share a preference for low-complexity regularities, without implying that all neural networks implement a universal Solomonoff prior (Goldblum et al., 2023).
2. Predictive memory, fairness, and the costs of coarse representations
Simplicity can arise through aggregation: distinguishable individuals are grouped into coarse cells and assigned a common prediction. “Simplicity Creates Inequity” studies an 5-approximator that partitions applicants into cells 6 and assigns each cell its average productivity,
7
A simple approximator restricts cells to cubes formed by fixing some attributes and ignoring others. This formalizes variable selection, decision trees, rule lists, and conjunctive categories.
The population contains an advantaged group 8 and disadvantaged group 9, with group membership having no causal effect on productivity:
0
Disadvantage is instead represented by a likelihood-ratio ordering: higher-productivity feature vectors are relatively more common among 1 than among 2. Under this condition, for a set 3 containing at least two productivity values,
4
The distinction is essential: a difference in aggregate group means alone is insufficient, and the analysis notes counterexamples based on Simpson’s paradox (Kleinberg et al., 2018).
Applicants are ranked by the approximator and the top fraction 5 is admitted. Efficiency is the average true productivity of admitted applicants,
6
while equity is the fraction of admitted applicants belonging to 7,
8
The central theorem states that every nontrivial simple 9-approximator is strictly improvable: there exists a more detailed approximator 0 satisfying
1
for every admission rate, with both inequalities strict at some rate. The more general theorem applies to every graded approximator, a class containing all simple approximators.
The mechanism is hidden heterogeneity. If a mixed cell contains a high-productivity 2 row, extracting it can create a cell with both higher productivity and greater 3-representation. Alternatively, extracting low-productivity 4 applicants can remove individuals who dilute efficiency and equity. If nontrivial cells are group-separable, the disadvantage condition can imply that the same feature vector is assigned a higher score in group 5 than in group 6, permitting a refinement that improves both cumulative outcomes.
A second theorem concerns group-agnostic simplification. If identical feature vectors are assigned to the same cell across groups, splitting a nontrivial cell into 7 and 8 yields
9
The split improves efficiency by ranking the higher-average-0 1 subcell first, but lowers equity because that subcell has zero 2 representation. Thus, after information is suppressed, group membership becomes predictive of the omitted features. Simplification can create an incentive for statistical discrimination even when group membership was irrelevant before aggregation.
The conclusions are conditional. They rely on cell-and-average approximation, cube or graded structure, strong likelihood-ratio disadvantage, genericity, a ranking-and-admission objective, and equity defined as the fraction of admitted 3 applicants. They do not establish that every interpretable, transparent, or low-complexity model is unfair. The result instead shows that interpretability achieved by coarse aggregation can be Pareto-dominated by a more detailed rule in the formal population-ranking problem.
3. Neural-network feature and representation biases
In neural networks, simplicity bias is commonly operationalized as preferential reliance on features with simpler decision boundaries. “The Pitfalls of Simplicity Bias in Neural Networks” defines feature simplicity by the minimum number of linear pieces needed to achieve optimal classification accuracy using that feature. A linear feature requires one threshold, whereas a 4-slab feature requires a decision boundary with 5 linear pieces. This definition is functional rather than semantic: it does not equate simplicity with interpretability, robustness, causality, or human-level abstraction (Shah et al., 2020).
The paper constructs datasets containing multiple independently predictive features with different simplicity levels. In LMS-6, one linear coordinate coexists with 7-slab coordinates; in noisy-LMS-8, the linear feature is noisy while the slab features are noiseless; MS-(5,7) contrasts 5-slab and 7-slab features; and MNIST–CIFAR concatenates an MNIST image with a CIFAR-10 image. Models are evaluated by randomizing feature subsets. If randomizing 9 reduces performance to chance while randomizing 0 leaves performance unchanged, the model depends critically on 1 and is invariant to 2.
Across fully connected networks, CNNs, and GRUs, LMS-5 produced approximately 3 randomized AUC when the simple linear feature was randomized and approximately 4 when the complex 5-slab features were randomized. The same pattern persisted with as many as 249 complex features. On MNIST–CIFAR, modern image architectures retained approximately 5 AUC when CIFAR was randomized but fell to approximately 6–7 when MNIST was randomized. The models therefore ignored predictive complex features despite sufficient capacity and training data to learn them.
The theoretical LSN result analyzes a one-hidden-layer ReLU network with one perfectly predictive linear coordinate, one perfectly predictive 3-slab coordinate, and Gaussian noise coordinates. Under the stated sample-size, width, initialization, and optimization conditions, after 8 iterations the learned weights associated with the linear coordinate are roughly 9 larger than those associated with the slab coordinate. The network reaches test error at most 0 while essentially ignoring the more complex predictive feature. The proof attributes this to a persistent linear-coordinate gradient and cancellation in the slab-coordinate gradient.
Extreme simplicity bias can explain non-robustness. If a model relies exclusively on one simple feature, a distribution shift that changes that feature can destroy performance while leaving other predictive features intact. In LMS-5, a classifier using many complex features can have margin of order 1, whereas the simplicity-biased network has effective margin 2. Universal adversarial perturbations concentrated approximately 3 of their 4 energy on the linear feature in LMS-5, approximately 5 on the 5-slab feature in MS-(5,7), and approximately 6 on MNIST pixels in MNIST–CIFAR.
Simplicity bias can also harm in-distribution generalization. In noisy-LMS, one noisy linear feature has 7 predictive accuracy while 49 noiseless 7-slab features have 8 predictive accuracy. Standard training continues to rely on the noisy linear feature, yielding reported test accuracies around 9–0 and approximately 1–2 accuracy after randomizing the 7-slab features. Thus, simplicity is not synonymous with predictive quality.
Several common remedies do not necessarily remove the bias. Ensembles help when independently trained models choose different equally simple features, as in MS-5, but fail when all models select the same noisy feature in noisy-LMS. Adversarial training can improve robust accuracy while leaving feature reliance unchanged: on MNIST–CIFAR, adversarially trained models achieved high robust accuracy under the specified attack but remained near chance when CIFAR was randomized. These results show that robustness to a particular perturbation need not imply use of richer or more semantically stable features.
A related formalization is low-dimensional simplicity bias. A model 3 exhibits LD-SB if there exists a low-rank projection 4 such that
5
while an independently trained model using 6 remains accurate. In the rich regime of a one-hidden-layer network, the principal theorem identifies a rank-one projection onto a linearly separable coordinate, even when orthogonal coordinates contain perfect classifiers. In the lazy regime, the first-layer weights can remain high rank while the function still depends on a low-dimensional projection (Morwani et al., 2023).
This distinction separates parameter simplicity from functional simplicity. A model may have high-rank weights but a low-dimensional decision function. Projection-mixing accuracy, logit changes, and orthogonal-subspace retraining provide empirical diagnostics. On frozen ResNet-50 representations, rank-one or low-rank projections explained predictions on binary Imagenette, Imagenette, Waterbirds, and MNIST–CIFAR while the complementary subspaces remained predictive.
Simplicity bias can also be hierarchical rather than exclusive. Under imbalanced label coupling, a network may first use a simple feature to select a coarse label group and then use a complex feature to distinguish labels within that group. In MNIST–CIFAR experiments, digit identity selected groups such as automobile/cat or dog/truck, while CIFAR information remained visible in structured confusions: cats were often predicted as dogs and trucks as automobiles. The resulting behavior resembles an implicit decision tree (Du, 2023).
The hierarchy can extend across three levels: patch features select a broad group, MNIST digits refine it, and CIFAR images supply the final class distinction. These results challenge the binary characterization “the network uses” versus “does not use” a feature. A complex feature can remain represented but be consulted only conditionally, after a simpler feature has constrained the prediction space.
4. Geometric, spectral, and architectural formulations
Several research programs define simplicity through the geometry of learned representations.
The low-rank simplicity-bias framework measures effective rank through spectral entropy. For singular values 7 normalized as
8
effective rank is
9
For a representation matrix 0, the paper constructs a cosine-kernel Gram matrix and measures its effective rank. A low-rank embedding means that examples are organized along a small number of dominant representational directions. In deep linear networks, increasing depth can preserve the represented function class while changing the parameterization. Products of random matrices increasingly concentrate their singular-value spectra, motivating the conjecture that deeper parameterizations contain a greater proportion of low-effective-rank functions (Huh et al., 2021).
The empirical evidence includes random initialization, least-squares tasks of different target ranks, noisy regression, optimizer comparisons, and linear over-parameterization of nonlinear networks. Deeper models tend to produce lower-effective-rank representations and often generalize better on low-rank or noisy tasks. They can underfit high-rank targets, showing that low-rank simplicity is task-dependent. Inserting additional linear factors into existing layers improved CIFAR-10 and CIFAR-100 performance, with approximately 1 expansion providing the best reported CIFAR results, while gains on ImageNet diminished with increasing baseline depth.
A related sharpness-based result studies a two-layer network with fixed second-layer weights and squared loss. On the zero-loss manifold, the trace of the Hessian is
2
Under positive-coherence data and activation assumptions including 3 and 4, all first-order stationary points of the trace-Hessian objective restricted to the interpolation manifold are global minimizers. They satisfy
5
for every neuron and training example. Therefore every neuron produces the same response, and the activation matrix
6
has rank one. Label-noise SGD is shown, in the small-step limit, to approach ordinary gradient flow until interpolation and then negative Riemannian gradient flow of the trace-Hessian objective on the zero-loss manifold. In this specialized setting, sharpness minimization yields a rank-one shared-feature representation (Gatmiry et al., 2024).
The result does not establish that flatness and simplicity are generally equivalent. It depends on the network, activation, loss, high-dimensional linearly independent data, fixed output weights, and trace-Hessian definition of sharpness. The rank-one statement concerns activations on the training data, not necessarily the ambient weight matrix.
Small initialization supplies another mechanism. For two-layer networks trained by gradient flow, early dynamics can decouple across neurons because the network output is quadratic in parameter scale and the logistic-loss derivative is approximately constant near zero. Neurons evolve according to a data-dependent function 7 independent of hidden width. Prominent neurons cluster around global extrema of 8, while other neurons remain smaller. In the middle stage, the network is approximately an embedding of a much smaller network. For XOR-like data, late training can preserve only four directions, intensifying the simplicity bias. The analysis suggests that intermediate-stage representations may transfer better OOD than fully converged representations, although the empirical conclusion is stated as a hypothesis rather than a universal theorem (Tsoy et al., 2024).
Transformer studies use different notions. Token-wise sensitivity is the expected fraction of token replacements that change the prediction:
9
For vision, token replacement is implemented by additive Gaussian noise on image patches; for language, token embeddings are layer-normalized before Gaussian perturbation. Transformers showed lower sensitivity than MLPs, CNNs, and LSTMs across the reported vision and language experiments. On Fashion-MNIST, ViT-small had sensitivity 00 versus 01 for the CNN; on CIFAR-10, ViT-small had 02 versus 03 for ResNet-18; and on MRPC, RoBERTa had sensitivity 04 versus 05 for the LSTM (Vasudeva et al., 2024).
In an infinite-width Boolean linear-attention setting, transformer kernels have Fourier eigenvalues ordered by degree within even and odd sectors:
06
This provides a weak spectral-bias explanation for learning lower-degree, lower-sensitivity functions earlier. The theorem does not apply directly to finite-width softmax transformers, and sensitivity depends on tokenization, perturbation distribution, patch size, and noise scale.
A distinct distributional formulation studies many-body token interactions. Factored-attention models with quadratic nonlinearities can express interaction orders growing as
07
with depth 08. In synthetic four-body distributions, one-layer models learned the best two-body approximation, deeper models subsequently learned three-body structure, and still deeper models recovered the full four-body distribution. In WikiText-103 and TinyStories, BERT and GPT models performed well on low-order clones early in training, after which low-order clone loss saturated while high-order clone loss continued to decline. This defines simplicity as the order of statistical interactions rather than parameter count, sensitivity, or geometric rank (Rende et al., 2024).
These formulations are related but nonidentical. Low effective rank, low sensitivity, low Fourier degree, low DNF length, and low interaction order can correlate in particular datasets while ranking functions differently in general.
5. Mitigation, alternative inductive biases, and model selection
Because simplicity bias can favor spurious or incomplete features, several methods aim to diversify or redirect learning.
Training a diverse set of models can expose solutions missed by ordinary SGD. One method penalizes alignment between input gradients of classifier heads:
09
The ERM term preserves predictive performance while the gradient penalty discourages identical local sensitivities. The method is not primarily an ensemble: it retains a collection of distinct predictors and uses independent OOD information to select one. On multi-dataset collages, standard independently trained models relied almost exclusively on MNIST, whereas the gradient-diversity method produced models using MNIST, SVHN, Fashion-MNIST, and CIFAR. It improved the best average accuracy from 10 for the independent collection to 11 for a 96-model collection. Improvements were also reported on Biased Activity Recognition and PACS (Teney et al., 2021).
OrthoP provides another diversity mechanism. It identifies dominant projection directions of one model, removes them, trains a second model on the orthogonal complement, and averages logits. On binary Imagenette, Waterbirds, and MNIST–CIFAR, OrthoP produced greater mistake diversity and lower class-conditioned logit correlation than independent training, and ensembles were more robust to Gaussian representation noise (Morwani et al., 2023).
SPARE exploits simplicity bias diagnostically. Early outputs separate majority and minority groups when spurious features have lower noise-to-signal ratio than core features. SPARE clusters examples within each class after a few training epochs and samples inversely to cluster size. It is designed to identify spurious groups even when majority examples are classified correctly, unlike methods based primarily on early misclassification. Reported worst-group accuracies were 12 on Colored MNIST, 13 on Waterbirds, 14 on CelebA, and 15 on UrbanCars. The method does not guarantee that clusters correspond to causal attributes, but it turns early simplicity-biased outputs into a diagnostic signal (Yang et al., 2023).
Long-tailed recognition introduces a class-frequency interaction. Tail classes receive fewer examples and therefore fewer opportunities to learn complex patterns. In MNIST-CIFAR-LT, supervised models relied heavily on the digit cue and neglected the CIFAR object, with the degradation especially severe for tail classes. Triple-Level Self-Supervised Learning combines holistic, partial, and augmented-level objectives. It improves representation diversity and reported ImageNet-LT performance to 16 overall, with 17 on few-shot classes. On MNIST-CIFAR-LT, CIFAR-only tail accuracy improved by 18 points at imbalance ratio 19 (Wei et al., 2023).
Activation functions provide a direct means of changing architectural inductive bias. A meta-learning procedure represents activations as piecewise-linear splines with approximately 50 control points over 20 and optimizes them through episodic validation performance. ReLU-like activations were nearly optimal on image classification, while higher-variation, sharper, feature-specific, or periodic activations improved image regression, tabular classification, shortcut-learning tasks, and algorithmic grokking tasks. Learned input activations produced axis-aligned boundaries resembling boosted trees on tabular data, and sine-like activations improved modular arithmetic. Across 22 algorithmic tasks, reported average accuracies were 21 for ReLU, 22 for task-specific learned activations, and 23 for one activation learned jointly across tasks (Teney et al., 13 Mar 2025).
These results qualify the common assumption that simplicity is universally beneficial. ReLU-induced simplicity was close to optimal for the image-classification tasks studied, but higher-complexity inductive biases were preferable for regression, tabular data, shortcut-prone environments, and algorithmic tasks. The appropriate inductive bias depends on the target’s structure rather than on a universal complexity ordering.
The Boolean DFCN framework makes this dependence explicit. Its hidden units represent conjunctions, and the network is in one-to-one correspondence with DNF formulas modulo hidden-unit and clause permutations. The principal complexity measure is minimum DNF literal count,
24
Uniform ternary initialization induces a prior 25 proportional to the fraction of discrete parameter configurations implementing 26. Weight decay multiplies this prior by approximately
27
Constants and sparse DNF functions receive high prior mass, while parity and random Boolean functions receive very low mass. Weight decay can improve generalization on simple and moderate-complexity targets but remains insufficient for high-complexity parity. The framework demonstrates that “simple” means simple relative to the architecture’s representation: DNF length is not equivalent to truth-table entropy, Fourier degree, Lempel–Ziv complexity, or Kolmogorov complexity (Mingard et al., 29 May 2025).
6. Open questions, controversies, and scope
A central controversy is whether simplicity bias should be interpreted as a universal explanation of generalization. The evidence supports several conditional claims: neural networks often favor simple or low-dimensional solutions; this can improve generalization when the target structure is low-dimensional or stable; and the same preference can cause shortcut reliance, poor robustness, unfair aggregation, or underfitting of complex targets. It does not support the stronger claim that simple solutions are always preferable.
A second issue concerns the meaning of “simple.” The literature includes at least the following non-equivalent measures:
| Simplicity notion | Primary object | Representative measure |
|---|---|---|
| Predictive memory | Internal predictive state | 28, 29 |
| Feature geometry | Representation spectrum | Effective rank 30 |
| Feature complexity | Decision boundary | Number of linear pieces |
| Functional dependence | Input projection | Rank of 31 in LD-SB |
| Local variation | Token or input perturbations | Sensitivity 32 |
| Algorithmic description | Individual sequence or dataset | 33, 34 |
| Logical structure | Boolean formula | 35 |
| Distributional structure | Many-body interactions | Interaction order |
A model can be low rank but semantically incomplete, low sensitivity but insensitive to useful information, algorithmically simple but geometrically irregular, or low DNF complexity but high truth-table entropy. Comparisons across papers therefore require specifying the representation, metric, data distribution, and optimization regime.
Theoretical results are generally restricted. Several use two-layer networks, infinite width, gradient flow, small balanced initialization, linearly independent data, specially constructed distributions, smoothed activations, or asymptotic random-matrix assumptions. Others provide primarily empirical evidence on synthetic datasets, frozen backbones, controlled image collages, or selected corruption distributions. Results obtained with trace-Hessian sharpness do not automatically transfer to SAM, largest-eigenvalue sharpness, ordinary SGD, ReLU networks, or deep architectures. Similarly, low sensitivity measured under Gaussian patch noise does not imply robustness to adaptive adversarial attacks or semantic perturbations.
A recurring methodological issue is causal interpretation. Feature randomization can show reliance but not necessarily explain why a feature was selected. Gradient diversity can generate distinct predictors without guaranteeing that one is causal. Activation visualizations and Network Dissection can reveal functional correlates without proving that those features cause robustness. Clone datasets can isolate interaction orders while remaining dependent on the generator architecture and imperfect sampling. SPARE can identify output clusters without automatically assigning them a human-interpretable spurious attribute.
The principal open problems include characterizing optimal quantum predictive representations; extending computational mechanics to higher-dimensional systems; determining when quantum and classical simplicity orderings reverse; developing finite-width theories of low-rank bias; formalizing simplicity bias for deep nonlinear architectures; understanding the relationship between sharpness, flatness, and function complexity; identifying when intermediate representations are preferable for OOD transfer; and establishing theory for self-supervised and transformer training beyond idealized kernels.
Other open directions concern intervention. Existing methods can diversify models, alter activations, rebalance clusters, mask attended regions, or train on orthogonal subspaces, but none guarantees recovery of causal features in all settings. A general solution would need to determine when simplicity is aligned with the intended task and when it is a shortcut. This is fundamentally a model-selection and data-design problem: i.i.d. observations may not identify which of several equally predictive explanations will remain valid under distribution shift.
Simplicity bias is therefore best understood as a family of measurable, framework-dependent inductive preferences rather than a single law. Its benefits arise when the favored representation matches the data-generating structure; its harms arise when simple features are spurious, noisy, inequitable, or insufficiently expressive. The central methodological implication is to measure the specific simplicity being induced—memory, rank, sensitivity, interaction order, logical length, or feature dependence—and to evaluate it against robustness, fairness, compositionality, and out-of-distribution objectives rather than treating simplicity itself as an unconditional virtue.