Publications
Preprints
Mind the Spike: Mechanisms and Brittleness of Visual Massive Activations in Large Vision–Language Models
Preprint — under review 2026
We study how visual activation spikes form in large vision–language models, their sensitivity to image perturbations, and an intervention that suppresses them.
Citation
Ngnawé, J., Pequignot, Y., Sahoo, S., Gagné, C., Precioso, F., & Koyejo, S. (2026). Mind the Spike: Mechanisms and Brittleness of Visual Massive Activations in Large Vision–Language Models. arXiv preprint arXiv:2609.32808.
A Stochastic–Geometric Theory of Scaling Laws in Grokking
Preprint — under review 2026
A theory connecting the geometry of optimization trajectories to delayed generalization, with grokking scaling laws for learning rate, batch size, and regularization.
Citation
Luo, R., Gagné, C., Ngnawé, J., Ullah, I., & Morrissey, K. (2026). A Stochastic–Geometric Theory of Scaling Laws in Grokking. arXiv preprint arXiv:2606.30388.
Conference Papers
RobustGenBench: A Benchmark for Robust Generalization to Adversarial and Common Perturbations, with Applications to Vision and Vision-Enabled Large Language Models
NeurIPS 2026, Evaluations and Datasets Track 2026
A benchmark for evaluating robust generalization to adversarial and common perturbations in vision models and vision-enabled large language models.
Citation
Heuillet, M., Ngnawé, J., Bhagwatkar, R., Pequignot, Y., Larouche, A., Gagné, C., Rish, I., Ahmad, O., & Durand, A. (2026). RobustGenBench: A Benchmark for Robust Generalization to Adversarial and Common Perturbations, with Applications to Vision and Vision-Enabled Large Language Models. NeurIPS 2026, Evaluations and Datasets Track.
Robust Fine-Tuning from Non-Robust Pretrained Models: Mitigating Suboptimal Transfer With Epsilon-Scheduling
ICLR 2026 2026
We study robust fine-tuning (RFT) of non-robust pretrained models and show that robust objectives cause suboptimal transfer. We propose Epsilon-Scheduling, which enables optimal transfer and improves expected robustness.
Citation
Ngnawé, J., Heuillet, M., Sahoo, S., Pequignot, Y., Ahmad, O., Durand, A., Precioso, F., Gagné, C. (2026). Robust Fine-Tuning from Non-Robust Pretrained Models: Mitigating Suboptimal Transfer With Epsilon-Scheduling. The Fourteenth International Conference on Learning Representations (ICLR 2026).
A Layer Selection Approach to Test Time Adaptation
AAAI 2025 2025
A gradient alignement method for layer selection in Test Time Adaptation.
Citation
Sahoo, S., ElAraby, M., Ngnawe, J., Pequignot, Y. B., Precioso, F., & Gagné, C. (2025, April). A layer selection approach to test time adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 39, No. 19, pp. 20237-20245).
Detecting Brittle Decisions for Free: Leveraging Margin Consistency in Deep Robust Classifiers.
NeurIPS 2024 2024
A novel property of deep robust classifiers that allows to use the logit margin as a proxy score for input margin and efficiently detect non-robust samples, vulnerable to adversarial attacks.
Citation
Ngnawé, J., Sahoo, S., Pequignot, Y., Precioso, F., & Gagné, C. (2024). Detecting Brittle Decisions for Free: Leveraging Margin Consistency in Deep Robust Classifiers. The Thirty-eighth Annual Conference on Neural Information Processing Systems.
Workshop Papers
Margin-based Progress Measures for Grokking
NeurIPS 2026 Workshop — Neural Network Artifacts as a New Data Modality 2026
We track instance margins (logit margin and distance to the decision boundary) as progress measures for grokking: test margins stay low during memorization and rise with delayed generalization.
Citation
Ngnawé, J., & Tikeng Notsawo, P. J. (2026). Margin-based Progress Measures for Grokking. NeurIPS 2026 Workshop on Neural Network Artifacts as a New Data Modality.
A Guide to Robust Generalization: The Impact of Architecture, Pre-training, and Optimization Strategy
NeurIPS 2025 Reliable ML Workshop 2025
This work presents the most comprehensive benchmark of robust fine-tuning to date, revealing how architecture, pretraining, and adaptation choices impact robust generalization across diverse datasets, perturbations, and training protocols.
Citation
Heuillet, M., Bhagwatkar, R., Ngnawé. J., Pequignot, Y., Larouche, A., Gagné, C., Rish, I., Ahmad, O., Durand, A. (2025). A Guide to Robust Generalization: The Impact of Architecture, Pre-training, and Optimization Strategy. arXiv preprint arXiv:2508.14079.
Robustmix: Improving Robustness by Regularizing the Frequency Bias of Deep Nets
NeurIPS 2022 Workshop on Distribution Shifts 2022
An extension of mixup in the frequency domain that regularizes the deep nets for robustness to common corruptions.
Citation
Ngnawe, J., NJIFON, M. A., Heek, J., and Dauphin, Y. Robustmix: Improving robustness by regularizing the frequency bias of deep nets. In NeurIPS 2022 Workshop on Distribution Shifts: Connecting Methods and Applications, 2022. URL https://openreview.net/forum?id=Na64z0YpOx.