We’re back after the summer break. In the fourth session, we’ll discuss “Raven: High-Recall Sequence Modeling with Sparse Memory Routing” by Afzal et al. [1, 2, 3]. While the Tuebingen AI Center team will host the discussion, we are proud and thankful to announce that the paper’s first author, Arshia Afzal, will join us for this installment. This 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.

LLMs’ arguably biggest strengths come from their ability to generate highly tailored output from their input sequence which is typically referred to as their context. In the last session, we have seen how state space model LLMs like Mamba-3 use their latent state to store this context while “mainline” LLMs use a sliding window if the input length becomes too large. Raven analyzes the different approaches under a common mathematical framework and proposes a new context management strategy based on a learned update mapping. We’ll discuss how they’ve turned this insight into a language model and what other challenges they had to overcome during training and evaluation.

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

[1] https://arxiv.org/abs/2607.25357 
[2] https://arshiaafzal.github.io/blog/2026/raven-part1/
[3]https://arshiaafzal.github.io/blog/2026/raven-part2/