Carlos Baquero
@xmal
Professor at FEUP University of Porto and researcher at INESC TEC. Distributed Systems and Data. Co-creator of CRDTs. Founder eurotux.com. Still searching for unknown unknowns. (🦋 July 2023) Web:
I start a sabbatical research year this September, will be doing contacts for short visits (still open to invites). As automation of reasoning expands, the contributions expected from humans will also be more ambitious. August, calm before the storm.
When multiple agents reason in parallel it is easy to merge their textual outputs, but very hard to merge their, richer, latent abstractions. We provide a first solution to this problem that also delivers CRDT properties.
Waiting for arXiv green light, and then likely a JMLR submission. First steps in bringing CRDTs to Latent Communication.
Vim nos dois últimos fim de semana ao Gerês e está surpreendentemente vazio. É de aproveitar.
Cache Merging as a Convergent Replicated State for Multi-Agent Latent Reasoning. Work with Luís Brito from IPVC.
Can we merge latent spaces as we merge CRDT replicas? It works for two replicas with a CRDT-like design, and we can do federated querying without exposing the query to the LLMs that process the source data. Abstract and link in thread.
We live in exciting times. Auto-research has been automated. Fable been running experiments all night.
Eating sardines🐟in summer is a tasty and healthy Portuguese tradition. Just an excuse to show our recent preprint on arXiv. arxiv.org/pdf/2605.30167
I wrote this in April, before Claude Code and Codex. Now I am changing my mind: The Carpenter can indeed just ask for a nice table.
Rateless Bloom Filters + Rateless Invertible Bloom Lookup Tables looks to be the perfect combination so far. Looking forward to seeing this help out with practical applications. Once the ArXiv link is up, we will also share the codebase link for reproducibility.
Much less overhead than competing solutions, unless sets are already very similar. And, unlike PinSketch, the computation cost is fine. It also shows the PinSketch lower bound was specific to sets of fixed sized elements.
Rateless Bloom Filters: Set Reconciliation for Divergent Replicas with Variable-Sized Elements. Preview of preprint waiting for ArXiv link.
Great attendance at DARE this year. This Erasmus+/ACM school is becoming a regular meeting point for the EU Distributed Data community, but we need more participation from the US. dare-lisbon.github.io/program/
Oh boy. I got myself again into this situation of preparing slides on the train for a talk at DARE 2025 tomorrow morning. All on track now, ou intended, but I slightly panicked two hours ago when started editing.