Eros Fanì

ML Researcher @ Gensyn
Building scalable, distributed models over uniquely decentralised and heterogeneous infrastructure

[gensyn @gensynai]

https://t.co/q9juIAYwTu

[Semih Kara @SemihSmile]

Multi-agent LLM systems are everywhere, and agents talk to each other: critiquing, correcting, giving feedback.

A design lever: what style of agent-to-agent communication works best?

At @gensynai, we designed one. New style, same agents, better results.

https://t.co/uYq1vdUjrC

[Eros Fanì @ErosFani]

@nlongcroft The repo is public and accessible

[Eros Fanì @ErosFani]

Look beyond one-size-fits-all LLMs with IR3DE

Introducing IR3DE: a router that selects the best expert LLM for any prompt.

No complex neural network or gradient-based training—just linear algebra and a linear layer achieving 98.4% in reasoning settings

https://t.co/acw7b1TiZ9

[Eros Fanì @ErosFani]

You just need to use token embeddings to construct ridge regression matrices, which can be done asynchronously. Then, you aggregate them and compute the optimal RMSE weights. The final weights of the token router are completely independent from the order you extract the statistics, and optimal in the sense of the RMSE objective.

Check out our paper and blog post for more information

https://t.co/VegS3zEA9J

https://t.co/acw7b1TiZ9

[gensyn @gensynai]

Look beyond one-size-fits-all LLMs with IR3DE

New research published by Gensyn's @ErosFani and @oguzer90

Read the blog post linked below and follow through to the published paper on @arxiv

[Eros Fanì @ErosFani]

Bonus: you can add or remove expert models without retraining the router from scratch.

Cheap, fast, and adaptable—useful as the list of available LLMs keeps growing.

[Eros Fanì @ErosFani]

Most expert routers need a full language model plus all your domain data pooled in one place to train.

IR3DE skips both. Domain stats are computed separately and then summed—ideal for privacy-constrained, decentralized settings.

[gensyn @gensynai]

At 11am ET join us as @ErosFani and @oguzer90 talk through the recently released research paper:

IR3DE: A Linear Router for Large Language Models

https://t.co/p13sKwgisp

[gensyn @gensynai]

https://t.co/P66GMCWHn3

[gensyn @gensynai]

Join @ErosFani tomorrow as he talk through the recent paper on DUME - Dynamic Upcycling of Mixture-of-Experts.

https://t.co/yMoYbqwBR7

[gensyn @gensynai]

$AI powers Gensyn

April 29

[gensyn @gensynai]

Gensyn team members are in attendance at two major conferences today - the International Conference on Learning Representations in Rio and EuroMLSys in Edinburgh

Find @oguzer90 and @ErosFani with papers at @iclr_conf and Nikolay Blagoev at @euromlsys

Details below👇

[gensyn @gensynai]

A reminder for those attending @iclr_conf on April 27, Gensyn team members will be on site with two papers being presented.

"Training-Free Dynamic Upcycling of Expert Language Models"

"Backdoor Attacks on Decentralised Post-Training"

Find details below and join them for more.

[Oguzhan Ersoy @oguzer90]

@ErosFani and I are at ICLR’26. If you’re here and interested in decentralised AI, let’s meet!

[gensyn @gensynai]

ICLR 2026, Rio de Janeiro, Apr 27 - Two Gensyn papers will be presented at @iclr_conf:

"Training-Free Dynamic Upcycling of Expert Language Models"

"Backdoor Attacks on Decentralised Post-Training"

Gensyn team members will be there and workshop details can be found below👇

[gensyn @gensynai]

"DUME provides a clean way to combine dense experts into a multidomain MoE – no compromises."

Newly published research by Gensyn's @ErosFani, introducing a training-free way to turn multiple domain experts into a single multi domain model.

https://t.co/zJkwwH2pAb

[Eros Fanì @ErosFani]

3/
Even better, DUME allows you to dynamically add experts in new domains or remove existing ones without retraining, and it works naturally in decentralized and privacy-sensitive settings, scaling intelligence in a modular, efficient, and composable way.

[Eros Fanì @ErosFani]

Stop multitask training. Just DUME.

Today, we introduce a training-free way to turn multiple domain experts into a single multi-domain model. DUME is training-free, scalable, and cheap.

https://t.co/MFdpsSO3fL

[Eros Fanì @ErosFani]

2/
Today's problems:

- Separate expert models, embedding: heavy memory + slow inference
- Model merging = destructive interference
- Multitask training = complex, expensive, unstable

DUME provides a clean way to combine dense experts into a multidomain MoE – no compromises.