mlsanityData to 7 Oct 2026

Mechanistic Mode Connectivity

arXiv 2211.08422 ↗First posted 15 Nov 2022cs.LG cs.CVcited by 50

Not on the map. Why?

Abstract (arXiv)

We study neural network loss landscapes through the lens of mode connectivity, the observation that minimizers of neural networks retrieved via training on a dataset are connected via simple paths of low loss. Specifically, we ask the following question: are minimizers that rely on different mechanisms for making their predictions connected via simple paths of low loss? We provide a definition of mechanistic similarity as shared invariances to input transformations and demonstrate that lack of linear connectivity between two models implies they use dissimilar mechanisms for making their predictions. Relevant to practice, this result helps us demonstrate that naive fine-tuning on a downstream dataset can fail to alter a model's mechanisms, e.g., fine-tuning can fail to eliminate a model's reliance on spurious attributes. Our analysis also motivates a method for targeted alteration of a model's mechanisms, named connectivity-based fine-tuning (CBFT), which we analyze using several synthetic datasets for the task of reducing a model's reliance on spurious attributes.

Key facts. Cited by 50 papers in the map (data through 7 Oct 2026). Builds on 107 of its 127 references that are in the map. First posted on arXiv 15 Nov 2022.

Builds on

107 of 127 references.

References · 107

  1. 2020Shortcut Learning in Deep Neural Networkscited by 1,928
  2. 2020The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalizationcited by 1,842
  3. 2019Invariant Risk Minimizationcited by 1,832
  4. 2019Do ImageNet Classifiers Generalize to ImageNet?cited by 1,507
  5. 2018ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustnesscited by 1,430
  6. 2022Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?cited by 1,413
  7. 2022Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timecited by 1,366
  8. 2018On the Spectral Bias of Neural Networkscited by 1,321
  9. 2016Deep CORAL: Correlation Alignment for Deep Domain Adaptationcited by 1,194
  10. 2018Averaging Weights Leads to Wider Optima and Better Generalizationcited by 1,194
  11. 2021Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gaugescited by 1,131
  12. 2018Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representationscited by 929
38 more
  1. 2019Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inferencecited by 864
  2. 2021Robust fine-tuning of zero-shot modelscited by 785
  3. 2020Out-of-Distribution Generalization via Risk Extrapolation (REx)cited by 737
  4. 2022Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distributioncited by 722
  5. 2018An Empirical Study of Example Forgetting during Deep Neural Network Learningcited by 669
  6. 2018Recognition in Terra Incognitacited by 655
  7. 2018Gradient Descent Provably Optimizes Over-parameterized Neural Networkscited by 655
  8. 2017The Implicit Bias of Gradient Descent on Separable Datacited by 649
  9. 2019Linear Mode Connectivity and the Lottery Ticket Hypothesiscited by 627
  10. 2018Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNscited by 585
  11. 2018Gradient Descent Finds Global Minima of Deep Neural Networkscited by 572
  12. 2019Fine-Grained Analysis of Optimization and Generalization for Overparameterized Two-Layer Neural Networkscited by 562
  13. 2019Variational Autoencoders and Nonlinear ICA: A Unifying Frameworkcited by 494
  14. 2021Fast Model Editing at Scalecited by 491
  15. 2020Underspecification Presents Challenges for Credibility in Modern Machine Learningcited by 488
  16. 2022Git Re-Basin: Merging Models modulo Permutation Symmetriescited by 460
  17. 2020Measuring Robustness to Natural Distribution Shifts in Image Classificationcited by 452
  18. 2022Last Layer Re-Training is Sufficient for Robustness to Spurious Correlationscited by 379
  19. 2018Rethinking ImageNet Pre-trainingcited by 376
  20. 2020What is being transferred in transfer learning?cited by 365
  21. 2020The Pitfalls of Simplicity Bias in Neural Networkscited by 349
  22. 2020Noise or Signal: The Role of Image Backgrounds in Object Recognitioncited by 342
  23. 2018Essentially No Barriers in Neural Network Energy Landscapecited by 328
  24. 2020An Investigation of Why Overparameterization Exacerbates Spurious Correlationscited by 310
  25. 2019Gradient Descent Maximizes the Margin of Homogeneous Neural Networkscited by 308
  26. 2021Self-Supervised Learning with Data Augmentations Provably Isolates Content from Stylecited by 308
  27. 2021The Role of Permutation Invariance in Linear Mode Connectivity of Neural Networkscited by 285
  28. 2018On the Optimization of Deep Networks: Implicit Acceleration by Overparameterizationcited by 280
  29. 2016Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICAcited by 270
  30. 2020Gradient Starvation: A Learning Proclivity in Neural Networkscited by 262
  31. 2018Implicit Bias of Gradient Descent on Linear Convolutional Networkscited by 251
  32. 2020Weakly-Supervised Disentanglement Without Compromisescited by 237
  33. 2016Deep Learning without Poor Local Minimacited by 233
  34. 2019Model Fusion via Optimal Transportcited by 217
  35. 2021Towards Causal Representation Learningcited by 207
  36. 2019SGD on Neural Networks Learns Functions of Increasing Complexitycited by 203
  37. 2018Deep learning generalizes because the parameter-function map is biased towards simple functionscited by 180
  38. 2018Algorithmic Regularization in Learning Deep Homogeneous Models: Layers are Automatically Balancedcited by 172

50 most cited of 107 All 107 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2022: 2; 2023: 14; 2024: 16; 2025: 17; 2026: 1.

50 papers.

Sort:

By area:

  1. 2022Editing Models with Task Arithmeticcited by 1,076
  2. 2024The Platonic Representation Hypothesiscited by 362
  3. 2024Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modificationscited by 220
  4. 2024Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMscited by 158
  5. 2023The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networkscited by 153
  6. 2024Eight Methods to Evaluate Robust Unlearning in LLMscited by 153
  7. 2023Mechanistically analyzing the effects of fine-tuning on procedurally defined taskscited by 110
  8. 2023Deep Model Fusion: A Surveycited by 89
  9. 2024Defending Against Unforeseen Failure Modes with Latent Adversarial Trainingcited by 82
  10. 2022ColD Fusion: Collaborative Descent for Distributed Multitask Finetuningcited by 52
  11. 2023Going Beyond Linear Mode Connectivity: The Layerwise Linear Feature Connectivitycited by 50
  12. 2024ICLR: In-Context Learning of Representationscited by 49
38 more
  1. 2023Merging by Matching Models in Task Parameter Subspacescited by 36
  2. 2023Quantification of Uncertainty with Adversarial Modelscited by 28
  3. 2023Proving Linear Mode Connectivity of Neural Networks via Optimal Transportcited by 24
  4. 2023Layer-wise Linear Mode Connectivitycited by 22
  5. 2023Compositional Capabilities of Autoregressive Transformers: A Study on Synthetic, Interpretable Taskscited by 21
  6. 2023A Closer Look at Model Adaptation using Feature Distortion and Simplicity Biascited by 20
  7. 2024C2M3: Cycle-Consistent Multi-Model Mergingcited by 20
  8. 2025Beyond the Permutation Symmetry of Transformers: The Role of Rotation for Model Fusioncited by 16
  9. 2023In-Context Learning Dynamics with Random Binary Sequencescited by 13
  10. 2025Unveiling the Basin-Like Loss Landscape in Large Language Modelscited by 13
  11. 2025From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularizationcited by 12
  12. 2025Understanding Mode Connectivity via Parameter Space Symmetrycited by 10
  13. 2023Project and Probe: Sample-Efficient Domain Adaptation by Interpolating Orthogonal Featurescited by 9
  14. 2024Towards an Understanding of Stepwise Inference in Transformers: A Synthetic Graph Navigation Modelcited by 9
  15. 2025The Butterfly Effect: Neural Network Training Trajectories Are Highly Sensitive to Initial Conditionscited by 9
  16. 2024Improving Model Fusion by Training-time Neuron Alignment with Fixed Neuron Anchorscited by 7
  17. 2025PEER pressure: Model-to-Model Regularization for Single Source Domain Generalizationcited by 7
  18. 2025Shared Parameter Subspaces and Cross-Task Linearity in Emergently Misaligned Behaviorcited by 7
  19. 2024Roadmap on Incentive Compatibility for AI Alignment and Governance in Sociotechnical Systemscited by 6
  20. 2024Do Deep Neural Network Solutions Form a Star Domain?cited by 6
  21. 2024Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defensecited by 6
  22. 2025Paths and Ambient Spaces in Neural Loss Landscapescited by 6
  23. 2023Confidence-Based Model Selection: When to Take Shortcuts for Subpopulation Shiftscited by 5
  24. 2024Input Space Mode Connectivity in Deep Neural Networkscited by 4
  25. 2024Foldable SuperNets: Scalable Merging of Transformers with Different Initializations and Taskscited by 4
  26. 2025Misalignment or misuse? The AGI alignment tradeoffcited by 4
  27. 2025Downgrade to Upgrade: Optimizer Simplification Enhances Robustness in LLM Unlearningcited by 4
  28. 2025Connecting Independently Trained Modes via Layer-Wise Connectivitycited by 3
  29. 2023Transferring Learning Trajectories of Neural Networkscited by 2
  30. 2025Evaluating Explanations: An Explanatory Virtues Framework for Mechanistic Interpretability -- The Strange Science Part I.iicited by 2
  31. 2025Evading Data Provenance in Deep Neural Networkscited by 2
  32. 2024Protecting against simultaneous data poisoning attackscited by 1
  33. 2024Exploring End-to-end Differentiable Neural Charged Particle Tracking -- A Loss Landscape Perspectivecited by 0
  34. 2025Find A Winning Sign: Sign Is All We Need to Win the Lotterycited by 0
  35. 2025Understanding Machine Unlearning Through the Lens of Mode Connectivitycited by 0
  36. 2025SAM Encoder Breach by Adversarial Simplicial Complex Triggers Downstream Model Failurescited by 0
  37. 2025Revisiting the Relation Between Robustness and Universalitycited by 0
  38. 2026Understanding Machine Unlearning Through the Lens of Mode Connectivitycited by 0