For vision transformers, we train linear probes on representations from individual tokens or on the representation averaged over all tokens, at the output of different transformer layers (each layer meaning a full transformer block including self-attention and MLP). Moreover, ResNets differ from ViTs in that the number of channels changes throughout the model, with fewer channels in the earlier layers. Wetrain alinear probe on each individual token and plot the average accuracy over the test set, in percent. Here we plot the results for each token a subset of layers in 3models: ViT-B/32 trained with aclassification token (CLS) or global average pooling (GAP), as well as a ResNet50. There are two main observations tobemade.

We thank the reviewers for their detailed and thoughtful comments. These are not new and have been presented thoroughly in the submitted paper. Our intention was not to challenge the momentum mechanism. Combining SwA V with a momentum encoder and/or a large memory bank are indeed interesting follow-ups. In Tab.5, we make a best effort fair comparison (same data augmentation, num.

Representations learned by pre-training a neural network on a large dataset are increasingly used successfully to perform a variety of downstream tasks. In this work, we take a closer look at how features are encoded in such pre-trained representations. We find that learned representations in a given layer exhibit a degree of diffuse redundancy, ie, any randomly chosen subset of neurons in the layer that is larger than a threshold size shares a large degree of similarity with the full layer and is able to perform similarly as the whole layer on a variety of downstream tasks. For example, a linear probe trained on $20\%$ of randomly picked neurons from the penultimate layer of a ResNet50 pre-trained on ImageNet1k achieves an accuracy within $5\%$ of a linear probe trained on the full layer of neurons for downstream CIFAR10 classification. We conduct experiments on different neural architectures (including CNNs and Transformers) pre-trained on both ImageNet1k and ImageNet21k and evaluate a variety of downstream tasks taken from the VTAB benchmark. We find that the loss \& dataset used during pre-training largely govern the degree of diffuse redundancy and the critical mass of neurons needed often depends on the downstream task, suggesting that there is a task-inherent redundancy-performance Pareto frontier. Our findings shed light on the nature of representations learned by pre-trained deep neural networks and suggest that entire layers might not be necessary to perform many downstream tasks. We investigate the potential for exploiting this redundancy to achieve efficient generalization for downstream tasks and also draw caution to certain possible unintended consequences.

Large Language Models (LLMs) often exhibit significant behavioral shifts when they perceive a change from a real-world deployment context to a controlled evaluation setting, a phenomenon known as "evaluation awareness." This discrepancy poses a critical challenge for AI alignment, as benchmark performance may not accurately reflect a model's true safety and honesty. In this work, we systematically quantify these behavioral changes by manipulating the perceived context of prompts. We introduce a methodology that uses a linear probe to score prompts on a continuous scale from "test-like" to "deploy-like" and leverage an LLM rewriting strategy to shift these prompts towards a more natural, deployment-style context while preserving the original task. Using this method, we achieved a 30% increase in the average probe score across a strategic role-playing dataset after rewriting. Evaluating a suite of state-of-the-art models on these original and rewritten prompts, we find that rewritten "deploy-like" prompts induce a significant and consistent shift in behavior. Across all models, we observed an average increase in honest responses of 5.26% and a corresponding average decrease in deceptive responses of 12.40%. Furthermore, refusal rates increased by an average of 6.38%, indicating heightened safety compliance. Our findings demonstrate that evaluation awareness is a quantifiable and manipulable factor that directly influences LLM behavior, revealing that models are more prone to unsafe or deceptive outputs in perceived test environments. This underscores the urgent need for more realistic evaluation frameworks to accurately gauge true model alignment before deployment.

In this work we study how explicit world-modeling objectives affect the internal representations and downstream capability of Transformers across different training stages. We use a controlled 2x2x2 Rubik's Cube and ask: (1) how does explicitly pretraining a world model affect the model's latent representations, and (2) how does world-model quality affect the model's performance after reinforcement learning post-training? We compare standard next-token prediction to two explicit world-modeling strategies -- (i) state-prediction pretraining and (ii) a joint state-prediction + next-token objective -- and assess task performance after Group Relative Policy Optimization (GRPO) is applied as post-training. We evaluate the representation quality with linear probes and causal interventions. We find that explicit world-modeling yields more linearly decodable and causally steerable state representations. More importantly, we find that improved state representations lead to higher gains for GRPO, especially on harder cube states. Our results indicate that sharpening state representations can improve the effectiveness of post-training for sequence-planning tasks.

The task of dataset distillation aims to find a small set of synthetic images such that training a model on them reproduces the performance of the same model trained on a much larger dataset of real samples. Existing distillation methods focus on synthesizing datasets that enable training randomly initialized models. In contrast, state-of-the-art vision approaches are increasingly building on large, pre-trained self-supervised models rather than training from scratch. In this paper, we investigate the problem of distilling datasets that enable us to optimally train linear probes on top of such large, pre-trained vision models. We introduce a method of dataset distillation for this task called Linear Gradient Matching that optimizes the synthetic images such that, when passed through a pre-trained feature extractor, they induce gradients in the linear classifier similar to those produced by the real data. Our method yields synthetic data that outperform all real-image baselines and, remarkably, generalize across pre-trained vision models, enabling us, for instance, to train a linear CLIP probe that performs competitively using a dataset distilled via a DINO backbone. Further, we show that our distilled datasets are exceptionally effective for fine-grained classification and provide a valuable tool for model interpretability, predicting, among other things, how similar two models' embedding spaces are under the platonic representation hypothesis or whether a model is sensitive to spurious correlations in adversarial datasets.