news
Talks, papers and other updates.
Attended Dagstuhl Seminar 26401, Divided We Stand: Security and Privacy of Collaborative Learning — a great week of exchanging ideas on the open challenges in the field.
Our paper with Eros Fanì, Training-Free Dynamic Upcycling of Expert Language Models, is accepted at NeurIPS 2026! We introduce DUME, a way to combine already-trained domain experts into one Mixture-of-Experts model without post-merge training.
OPEN-1B is out — a fully auditable LLM whose training steps can be replayed and verified bit for bit. Preprint · announcement.
We built IR3DE-AXL, a collective inference network prototype where participants can contribute expert models, use models served by others, or both.
Wrote up F-TIS on the Gensyn blog — running GRPO across heterogeneous models without assuming they are all the same.
Backdoor Attacks on Decentralised Post-Training is accepted at the ICLR 2026 Workshop on Principled Design for Trustworthy AI. Preprint.
New preprint: Hail to the Thief — attacks and defenses in decentralised GRPO. (Now in Findings of ACL 2026.)
Two preprints out of the Gensyn research team: NoLoCo, a low-communication training method that avoids all-reduce, and All is Not Lost, on recovering LLM training without checkpoints.