Results

The cross-dataset synthesis, then drill into each domain.

In-distribution accuracy is not a reliable guide to temporal robustness. Across the three datasets, how much a model degrades under temporal shift is governed by the strength of its inductive bias and by whether it relies on a frozen pretrained encoder.

Overview

Across the three datasets, the temporal robustness of a model is governed by how much its inductive bias improves in-distribution accuracy. The features an architecture extracts to fit the training period are specific to that period and are the first to degrade as the data drifts, so the architectures that are most accurate in distribution tend to degrade fastest.

The strength of the effect scales with the in-distribution accuracy gap that the inductive bias opens between architectures.

On Yearbook in-distribution accuracy across scratch architectures spans roughly % to %, with the convolutional models at the top of that range, and the most accurate models lose up to points under forward evaluation. On Amazon recurrence and attention open a smaller gap and a correspondingly smaller decay. On arXiv the task reduces to topic recognition, which the frozen RoBERTa embeddings already capture, so the inductive bias yields no accuracy advantage between architectures: all families reach to % macro-AUC, and every architecture decays by a similar small amount.

// 01  three domains, one structure

Each model is summarised by a temporal drift matrix: trained on all data up to a cutoff, evaluated on every period, with the in-distribution diagonal read against the off-diagonal decay. The three domains span a century of yearbook portraits, a decade of Amazon reviews and twenty-five years of arXiv abstracts, on three different metrics and two directions, yet the same structure appears in each.

Yearbook, accuracy.
Amazon, balanced MSE (lower is better).
arXiv, macro-AUC.

Each matrix is seed-averaged first (Yearbook: 5 distinct model seeds; arXiv/Amazon: 3), then shown as the cohort mean over that dataset's models, on its own metric and direction.

// 02  inductive bias and pretraining

Inductive bias

A stronger inductive bias extracts more discriminative, period-specific features, which raise in-distribution accuracy but decay fastest under temporal shift.

On Yearbook the convolutional and residual networks are the most accurate and the least robust. On Amazon the recurrent and Transformer networks show the same pattern. On arXiv, where the bias provides no accuracy advantage, the architectures are indistinguishable in both accuracy and decay.

Pretraining

A frozen pretrained encoder occupies a different regime, conceding in-distribution accuracy in exchange for stability.

On Yearbook and Amazon the frozen encoders are less accurate than the trained models on the training period but degrade more slowly, and DeBERTa-v3 is the most temporally stable model on Amazon, at decay.

Pretraining adjusts the accuracy-robustness balance independently of the architecture's inductive bias.

// 03  conclusion

In-distribution accuracy neither guarantees nor predicts temporal robustness. The strength of a model's inductive bias and whether it relies on a frozen pretrained encoder both shape where it falls on the accuracy-robustness trade-off. Under temporal shift, the model that performs best on held-out data from the training period is often the least robust over time, so architecture selection for deployment should account for the expected horizon between retraining.

// 04  limitations

  • Limited set of architectures. We compare the most common inductive biases, convolutional, recurrent and attention-based, against simple baselines, but many architectures remain untested, so the study is best read as preliminary and open to extension.
  • Coverage of datasets and domains. The study spans three datasets across two modalities, images and text. Extending it to further datasets, additional modalities, and other forms of distribution shift would show how broadly the observed pattern holds.
  • Text label space and time span. The text tasks use a narrow label space, a five-point rating for Amazon and seven categories for arXiv, and span only about a decade and a quarter of a century respectively, against a century for Yearbook. A richer label space and, above all, a longer time span should surface stronger language drift.
  • Frozen text representations. The text models train a head on top of fixed RoBERTa embeddings rather than learning representations end-to-end, since the examples available within each time slice are too few to train a text encoder from scratch. The text results therefore reflect a single frozen representation, and their absolute values should be read in that light.
  • Subsampled text corpora. The text experiments run on a subsample of each corpus rather than the full data and complete category set. Repeating them on the full data and all categories could surface stronger drift signals than the subsample reveals.
  • Descriptive, not mechanistic. We measure how robustness varies with inductive bias and pretraining, but do not isolate why.
  • No adaptive retraining. We do not use a dynamic retraining policy or dynamic per-training data selection: for each cutoff, models train on all historical data available up to that point. The results speak to inherent temporal robustness, not to continuous ML orchestration systems such as the ETH Zurich Modyn paper Modyn: Data-Centric Machine Learning Pipeline Orchestration, which studies triggering and data-selection policies under drift.

// 05  future work

  • Broader architectures and domains. Extend the comparison to more inductive biases and to additional datasets, including domains beyond images and other forms of drift.
  • Longer spans and full data. Lengthen the temporal span and replicate on the full corpora and richer label spaces, where language drift should be more pronounced.
  • Adaptation under budget. Pair the drift matrices with scheduled or drift-triggered retraining under explicit compute budgets, to measure not only decay but how cheaply it is undone.
  • From description to mechanism. Tie inductive bias and pretraining back to the specific cues that drift.