Research shouldn't require permission.

Can two language models trained on different modalities (e.g., image and text) communicate in latent space to unlock better visual reasoning?

Redundancy-aware merging proposes combining two AI models by injecting only the distinct, unique features of Model A into Model B while discarding shared information. It achieves this either geometrically (by projecting out shared components to keep the orthogonal residual) or statistically (by removing linearly predictable elements), assuming both models operate in compatible activation spaces.

Do an LLM’s neuron activations specialize by the phase of automated program repair (localization, reproduction, patching, testing), and can that structure be exploited to make repair agents cheaper?

Can two language models coordinate by exchanging their internal state directly instead of collapsing their thinking into text? If so, what are the implications, and how can such communication be extended over different modalities (e.g. image-text, text-image).

Neural networks power everything from ChatGPT to face recognition, but what actually happens inside one? In this workshop we open the black box and follow a real prediction through a real network, number by number: how a neuron multiplies and adds, why activation functions "bend the line," how the loss measures a mistake, and how backpropagation and gradient descent turn errors into learning. No AI or math background needed — if you can multiply and add, you can follow everything. We finish by reading a complete neural network written from scratch in C++, so you leave having seen every line of code behind the "magic."

Most structured data lives in relational databases, but AutoML still assumes a single flat table — so humans spend their effort flattening. This project surveys AutoML and relational deep learning, sketches a formalism for agents that explore schemas, enrich features and search models directly on the database, and asks whether copy-on-write database clones are the right substrate for the many experiment versions such agents create.