MUX selected as a spotlight at NeurIPS 2026
MUX: Continuous Reasoning via Multiplexed Tokens was accepted to NeurIPS 2026 as a spotlight paper.
News
All the news from the group.
Saka Research has a new website
Our group website is live, with pages for our research, team, publications, and news.
New preprint: Global Transport Couplings for Classifier-Free Guided Flows
A class-agnostic optimal-transport coupling for conditional flow models that consistently improves generation when combined with classifier-free guidance.
New preprint: To Learn is to Wander
We released Wander, a single graph foundation model for node classification, link prediction, and knowledge-graph reasoning built on random walks.
RelAgent accepted to NeurIPS 2026
Our paper RelAgent: LLM Agents as Data Scientists for Relational Learning was accepted to NeurIPS 2026.
New preprint: Thinking with Looped Flows
Looped flows train recurrent models with local denoising objectives, so that extra computation at inference time can help solve harder problems.
Talks on graph foundation models
İsmail gave talks and guest lectures on graph foundation models at the LOGML summer school, EPFL, the Wallenberg Advanced Scientific Forum, and Oxford.
New preprint: What are the Right Symmetries for Formal Theorem Proving?
A category-theoretic framework for the symmetries that LLM-based theorem provers should respect.
Looking for older posts? See the news archive or subscribe via RSS.