Emanuele

Research associate at @UniofOxford.
Co-founder and co-director of @idai_institute.
Previously: @UniofOxford (PhD), @polimi (bachelor+master).

[Samuele Marro @MarroSamuele]

Calling all multi-agent AI researchers, join us for the IDAI Workshop on Multi-Agent Safety and Security!
A one-day gathering at the Royal Society Of Arts on creating safe, secure and reliable multi-agent systems.
Featuring keynotes by @casdewitt (University of Oxford) and Georgios Piliouras (DeepMind).
Date: June 29th, 2026

Location: Royal Society of Arts, 8 John Adam Street, London WC2N 6EZ

We will also hold a poster session on new developments in multi-agent safety and security.
To present a poster, submit an abstract (up to 150 words) to [email protected]
Deadline: 24 June 2026, 23:59 AOE
The best poster will receive 500 USD of compute provided by the Institute for Decentralized AI.

Spots are limited. Join the waitlist here: https://t.co/7FkTBjJ6sI

More info: https://t.co/drblWEOkd4

Organized in collaboration with @Microsoft , @idai_institute , @ethereumfndn's dAI Team and supported by @cosmos_inst

[Emanuele @iperboreo_]

@noranta4 che figata la maglia :D

[Emanuele @iperboreo_]

@wire9000 @Dott_P_M @micheleboldrin nessuna di queste cose richiede di sapere scrivere bene in italiano/inglese, tantomeno serve il latino :D

ma di che stai parlando

[Emanuele @iperboreo_]

@carloalberto @openclaw Ho diverse perplessità in merito: bisogna vedere se la decentralizzazione di data, compute, oversight e consenso offre incentivi tecnologici, economici e politici sufficienti per creare un'alternativa al modello centralizzato.

[Emanuele @iperboreo_]

@callmeMizuko or 11^353

[Emanuele @iperboreo_]

@callmeMizuko 353^11?

[Samuele Marro @MarroSamuele]

Our new paper, Permission Manifests for Web Agents, is out on arXiv! It's the first paper of the Lightweight Agent Standards Working Group.

In a sentence: robots.txt for AI agents

https://t.co/wbEWjEUwmI

AI agents have no way of knowing what interactions are allowed on webpages, which means they often break TOSes. The typical solution is for websites to block all AI agents (see @Cloudflare). Agents then try to circumvent anti-AI blockers, and so on.

The solution to this arms race is a standardized, machine-readable document that specifies how an AI agent is allowed to interact with a webpage.

agent-permissions.json allows webpages to specify both fine-grained HTML rules (“don’t click this button”) and general guidelines (“when registering an account, use _bot at the end of your username”). It also supports specifying alternative MCP, A2A and OpenAPI endpoints for the webpage.

Just as robots.txt addressed the problem of specifying rules for crawlers, agent-permissions.json is the first step towards specifying rules for UI interactions. It is also designed to allow AI-friendly webpages to explicitly welcome interactions from agents, which boosts visibility

You can find the standard here: https://t.co/V4mudoxYfU

We also released a Python library (https://t.co/kfDgDfbbYn), a web tool to generate agent-permissions.json files for your website (https://t.co/KC7GWC7bri), and a Python integration demo (https://t.co/F4ZuJm1bjJ).

Huge thanks to:
@_achan96_
@XinxingRen
@lrhammond
Jesse Wright
@rickywanga42
@tizianopiccardi
@nfcampos
@TobinSouth
Jialin Yu
@alex_pentland
Philip Torr
@jiaxin_pei

[Emanuele @iperboreo_]

@Gozuyht ci sono due impostazioni sullo scholar. quella che ti aggiunge in automatico i paper (e quindi rischi di aggiungere roba di omonimi), e quella dove fai tutto te a mano.
se aggiungi in automatico sarebbe buona norma eliminare paper non tuoi.

[Emanuele @iperboreo_]

@jxmnop it would be interesting to see, on average and in practice, what's the \epsilon under which most of the internal representations collapse to balls.

I reckon this has been partially addressed in the "llms' hidden reps are mostly anysotrhopic", though.

[Emanuele @iperboreo_]

Read the article here! 👉 https://t.co/FIwDXNmrob

A huge thanks to the people who contributed to this work: @NanarNarame, @MarroSamuele, @jiemarinazhang, @ElizabethBlack, @michaelluck, @philiptorr, @wooldridgemike

[Emanuele @iperboreo_]

🚨 Are we misusing the term Multi-Agent Systems in LLM research? 🚨

Find out more in our latest preprint:

"Large Language Models Miss the Multi-agent Mark"

More info ⬇

[Emanuele @iperboreo_]

In other words, MAS has decades of research to offer, yet much of it is being sidelined.

We call for:
• Clearer terminology
• Better integration of established MAS principles
• A more rigorous approach to designing truly multi-agent LLM systems

⬇

[Emanuele @iperboreo_]

@paolodune @AleEquilibrium Il fatto che la mia gamba e un prosciutto siano entrambe carne non e' una buona analogia perche' mi morda il polpaccio.

Si chiama patto generazionale quello di cui parli

[Emanuele @iperboreo_]

@paolodune @AleEquilibrium In che senso e' uno schema Ponzi?
Qua le caratteristiche: https://t.co/EHzXYkVevP

Mi potresti fare l'analogia che non capisco?

[Samuele Marro @MarroSamuele]

You can join by taking part in the Discord server: https://t.co/vRwD1yA9PC
Website: https://t.co/9iDRXiCXyd
Looking forward to working with y'all!

[Samuele Marro @MarroSamuele]

Everyone's trying to create the perfect standard for agents. We follow the opposite approach: we pick small, specific problems and standardize those.
Glad to announce the creation of the Lightweight Standard Agent Working Group!

[Emanuele @iperboreo_]

@MatternJustus We explored a bottom-up approach to generate benchmarks with code.
We defined some naturalistic tasks whose completion is fully captured as a code.
https://t.co/YTT9TMAHC0

People at deepseek took the reverse approach, from existing benchmarks to code.
https://t.co/ZSNWexghrz

[Emanuele @iperboreo_]

@natolambert Interesting piece of work.
We have been working on that problem at Oxford for ~1.5 years.
Initial work: https://t.co/qPAoVxeUFB
Recent update - released one week before them :D - : https://t.co/YTT9TMAHC0

[Emanuele @iperboreo_]

@janleike Given the nature of the challenge (classifier on the prompt, in embedding space and in the output + broken semantic similarity of the answer, which has to be in PLAIN ENGLISH) t's enough to break one to prove your system unsafe (many did)
Constitutional classifier, LEL