mlsanityData to 14 Oct 2026

Mechanistic Evidence for Preserved-but-Misaligned Representations in Non-IID FedAvg

arXiv 2512.23043 ↗First posted 28 Dec 2025cs.LGcited by 0

On the map: Mechanistic interpretability → Sparse autoencoder training and evaluation

Abstract (arXiv)

Federated Averaging (FedAvg) often degrades under non-IID client data, but it remains unclear whether this degradation reflects the loss of client-learned representations or a failure to use representations that are still present. We study this question mechanistically in sparse client-trained vision models, using dense-model controls to test whether the observed effects depend on sparsity. Our analysis combines class-specific circuit discovery, linear probing of frozen representations, head-only finetuning, and sparse feature dictionaries. Across CNN and ResNet models on CIFAR-10 and Fashion-MNIST, severe label skew can drive some per-class accuracies near zero even when class-specific internal structure remains recoverable. Linear probes substantially outperform the aggregated classifier, head-only finetuning partially restores accuracy, and USAE transfer reveals a largely shared feature basis between IID and non-IID models. Together, these diagnostics suggest that, in our setting, non-IID FedAvg degradation is not fully explained by representational erasure; it also reflects misalignment between preserved internal structure and the final prediction pathway.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Sparse autoencoder training and evaluation, one of 13 directions in Mechanistic interpretability. Builds on 24 of its 32 references that are in the map. First posted on arXiv 28 Dec 2025.

Builds on

24 of 32 references.

Closest ancestors · 1

Same direction, newest first

  1. 2022Toy Models of Superposition21 Sep

Other references · 23

  1. 2016Communication-Efficient Learning of Deep Networks from Decentralized Datacited by 7,669
  2. 2019Advances and Open Problems in Federated Learningcited by 3,046
  3. 2018Federated Optimization in Heterogeneous Networkscited by 2,987
  4. 2013Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computationcited by 2,079
  5. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  6. 2022Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timecited by 1,396
  7. 2016Understanding intermediate layers using linear classifier probescited by 1,353
  8. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  9. 2022Editing Models with Task Arithmeticcited by 1,120
  10. 2018Federated Learning with Non-IID Datacited by 1,108
  11. 2020Tackling the Objective Inconsistency Problem in Heterogeneous Federated Optimizationcited by 906
  12. 2023TIES-Merging: Resolving Interference When Merging Modelscited by 818
11 more
  1. 2021Model-Contrastive Federated Learningcited by 766
  2. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  3. 2019Linear Mode Connectivity and the Lottery Ticket Hypothesiscited by 637
  4. 2021Merging Models with Fisher-Weighted Averagingcited by 621
  5. 2019Federated Learning with Personalization Layerscited by 541
  6. 2021Exploiting Shared Representations for Personalized Federated Learningcited by 541
  7. 2022Git Re-Basin: Merging Models modulo Permutation Symmetriescited by 471
  8. 2019The Non-IID Data Quagmire of Decentralized Machine Learningcited by 305
  9. 2024Have Faith in Faithfulness: Going Beyond Circuit Overlap When Finding Model Mechanismscited by 149
  10. 2023Stochastic Controlled Averaging for Federated Learning with Communication Compressioncited by 101
  11. 2025Weight-sparse transformers have interpretable circuitscited by 44

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Do Vision Models Truly Forget? New Findings from Representation-Level Certification of Visual Unlearning in Vertical Federated Learning
  2. 2025OvA-LP: A Simple and Efficient Framework for Federated Learning on Non-IID Data
  3. 2026FedSIR: Spectral Client Identification and Relabeling for Federated Learning with Noisy Labels
  4. 2025FedReFT: Federated Representation Fine-Tuning with All-But-Me Aggregation
  5. 2024Exploring Vacant Classes in Label-Skewed Federated Learning
  6. 2025Federated Deconfounding and Debiasing Learning for Out-of-Distribution Generalization
  7. 2025Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry
  8. 2022Federated Zero-Shot Learning for Visual Recognition