Gradient Flow Polarizes Softmax Outputs towards Low-Entropy Solutions.

Mark Rofin

Aditya Varre Mark Rofin Nicolas Flammarion Gradient Flow Polarizes Softmax Outputs towards Low-Entropy Solutions. 2026 March abs/2603.06248 CoRR https://doi.org/10.48550/arXiv.2603.06248 db/journals/corr/corr2603.html#abs-2603-06248 streams/journals/corr

Mark Rofin Jalal Naghiyev Michael Hahn 0001 Understanding the Emergence of Seemingly Useless Features in Next-Token Predictors. 2026 March abs/2603.14087 CoRR https://doi.org/10.48550/arXiv.2603.14087 db/journals/corr/corr2603.html#abs-2603-14087 streams/journals/corr

Mark Rofin Aditya Varre Nicolas Flammarion (How) Learning Rates Regulate Catastrophic Overtraining. 2026 April abs/2604.13627 CoRR https://doi.org/10.48550/arXiv.2604.13627 db/journals/corr/corr2604.html#abs-2604-13627 streams/journals/corr

Alireza Amiri Bavandpour Xinting Huang Mark Rofin Michael Hahn 0001 Lower Bounds for Chain-of-Thought Reasoning in Hard-Attention Transformers. 2025 ICML https://proceedings.mlr.press/v267/bavandpour25a.html https://openreview.net/forum?id=Oh9sG5ae2b conf/icml/2025 db/conf/icml/icml2025.html#BavandpourHR025

Alireza Amiri Bavandpour Xinting Huang Mark Rofin Michael Hahn 0001 Lower Bounds for Chain-of-Thought Reasoning in Hard-Attention Transformers. 2025 February abs/2502.02393 CoRR https://doi.org/10.48550/arXiv.2502.02393 db/journals/corr/corr2502.html#abs-2502-02393 streams/journals/corr

Michael Hahn 0001 Mark Rofin Why are Sensitive Functions Hard for Transformers? 14973-15008 2024 ACL (1) https://doi.org/10.18653/v1/2024.acl-long.800 https://aclanthology.org/2024.acl-long.800 https://www.wikidata.org/entity/Q131458627 conf/acl/2024-1 db/conf/acl/acl2024-1.html#0001R24

Michael Hahn 0001 Mark Rofin Why are Sensitive Functions Hard for Transformers? 2024 abs/2402.09963 CoRR https://doi.org/10.48550/arXiv.2402.09963 db/journals/corr/corr2402.html#abs-2402-09963

Mark Rofin Vladislav Mikhailov Mikhail Florinsky Andrey Kravchenko Tatiana Shavrina Elena Tutubalina Daniel Karabekyan Ekaterina Artemova Vote'n'Rank: Revision of Benchmarking with Social Choice Theory. 670-686 2023 EACL https://doi.org/10.18653/v1/2023.eacl-main.48 https://aclanthology.org/2023.eacl-main.48 conf/eacl/2023 db/conf/eacl/eacl2023.html#RofinMFKSTKA23

Mark Rofin Vladislav Mikhailov Mikhail Florinskiy Andrey Kravchenko Elena Tutubalina Tatiana Shavrina Daniel Karabekyan Ekaterina Artemova Vote'n'Rank: Revision of Benchmarking with Social Choice Theory. 2022 abs/2210.05769 CoRR https://doi.org/10.48550/arXiv.2210.05769 db/journals/corr/corr2210.html#abs-2210-05769

Mark Rofin Nikita Balagansky Daniil Gavrilov Linear Interpolation In Parameter Space is Good Enough for Fine-Tuned Language Models. 2022 abs/2211.12092 CoRR https://doi.org/10.48550/arXiv.2211.12092 db/journals/corr/corr2211.html#abs-2211-12092

Ekaterina Artemova Nikita Balagansky Alireza Amiri Bavandpour Nicolas Flammarion Mikhail FlorinskyMikhail Florinskiy Daniil Gavrilov Michael Hahn 0001 Xinting Huang Daniel Karabekyan Andrey Kravchenko Vladislav Mikhailov Jalal Naghiyev Tatiana Shavrina Elena Tutubalina Aditya Varre