The third instalment of the online journal club will feature the third instalment of the Mamba model series. We will discuss and dissect the ICLR 2026 paper “Mamba-3: Improved Sequence Modeling using State Space Principles” by Lahoti et al. [1]. Again, the Tübingen AI Center team will host and lead the discussion. The session is open to all and free of charge. We encourage you to read the paper before joining, but it’s not a strict requirement. If you have specific questions that you want to have discussed during the session, send them to minerva@tuebingen.ai.

Mamba-3 [1] is the third instalment of an alternate LLM model architecture [2, 3]. While the majority of transformer-based LLM architectures are structured as auto-regressive models, mapping sequences of length n to sequences of length n+1, models in the Mamba family capture the state of the model in a latent state, thus circumventing the linear memory and quadratic compute cost at inference time. We will have a close look what implications this has both in terms of computational requirements, but also in model performance and what future improvements might be expected from models in this family.

Looking forward to see many of you online and diving deep into the research trends.
 

[1] https://openreview.net/forum?id=HwCvaJOiCj
[2] https://openreview.net/forum?id=ztn8FCR1td
[3] https://openreview.net/forum?id=tEYskw1VY2