←
Indro Spinelli🤠CoRL258 posts
Indro Spinelli🤠CoRL avatar

Indro Spinelli🤠CoRL

@IndroSpinelli

Tenure Track Assistant Professor @SapienzaRoma | @ELLISforEurope member.

◷ Joined December 2011▧ 33 Photos and videos
Matteo Migliarini avatar
⟳ Indro Spinelli🤠CoRL retweeted
Matteo Migliarini@mamiglia7·Oct 26

I’m presenting “Quantifying Self-Preservation Bias in Large Language Models” today at #EMNLP2026 in Budapest! Instrumental convergence predicts AI will resist being shut down. We tested it: many LLMs argue against their own replacement, even by a better model. 🧵👇

Indro Spinelli🤠CoRL avatar
Indro Spinelli🤠CoRL@IndroSpinelli·Oct 8

Even arbitrary labels can divide LLM agents into factions. In our study, revealing peers’ model families cut success from 96% to 81%. Shuffling the labels shifted who grouped with whom. A small prompt detail with real costs for multi-agent cooperation. 👇

[1/9] We gave LLM agents one piece of information: which model family each peer came from. They split into factions. The catch: nearly every frontier model's system prompt already says which model it is. So today's multi-agent systems are already running this experiment. 🧐

0:58
Phillip Isola avatar
⟳ Indro Spinelli🤠CoRL retweeted
Phillip Isola@phillip_isola·Sep 28

Students sometimes ask me if it still makes sense, in this accelerating age, to pursue a PhD in AI. Perhaps counterintuitively, I think it's a great time to do so. I wrote up some thoughts on this here: web.mit.edu/phillipi/www/wri…

Paradigma avatar
⟳ Indro Spinelli🤠CoRL retweeted
Paradigma@paradigmainc·Sep 21

Introducing Limite 1B - Violetto. A model for high-frequency mathematical intelligence.

Post attachment
Indro Spinelli🤠CoRL avatar
Indro Spinelli🤠CoRL@IndroSpinelli·Sep 11

Sorry for the quality but Yann’s talk got me excited. Saturday at 3 PM, pass by poster #365 to chat about world models that do not generate pixels and steer your robot towards safety and successful actions!

Post attachment
Indro Spinelli🤠CoRL avatar
Indro Spinelli🤠CoRL@IndroSpinelli·Sep 5

With BiCICLe, an off-the-shelf text LLM coordinates two arms via multi-agent in-context learning. It beats supervised baselines and transfers straight to real-world tasks. Huge effort by @palm_alessio and the team at @TUDarmstadt. See you in Texas for #CoRL2026! 🤖

1/🧵 Can an off-the-shelf, text-only LLM control two robot arms from just a few demonstrations? That’s the question behind BiCICLe, our work on bimanual robot manipulation, now accepted to #CoRL2026 🤖🤖

0:21
Thomas Wolf avatar
⟳ Indro Spinelli🤠CoRL retweeted
Thomas Wolf@Thom_Wolf·Jul 3

One of the clearest arguments I've read for why openness matters. Worth 2 minutes of your time. @andykonwinski puts into words something many of us have been feeling: -- "Democracy is built on a profound skepticism of concentrated power. Open science shares this principle. Both are built on the idea that progress and legitimacy emerge from broad, distributed participation rather than concentrated, gated authority." "If our best scientists and engineers can only reach the frontier by joining a handful of secretive labs, we do not have an open research ecosystem. We do not have a truly competitive market. We have a system in which participation increasingly depends on the permission of a few individuals at a small number of private companies." "The challenge now is to build a new commons at the intersection of academia, industry, and the public interest. This research commons must be ambitious enough to matter. It will require frontier-scale compute, access to state-of-the-art models, operational support, public investment, and philanthropic capital. It will require companies willing to contribute to an ecosystem larger than themselves, so that they can continue to benefit from open research."

Andy Konwinski @andykonwinski Jul 2 Article Concentration of power in AI is a risk, not a solution The answer to concentration of power in AI is not openness at all costs, but a serious research commons with the resources to compete. Over the past few months, I've become increasingly uneasy about

Indro Spinelli🤠CoRL avatar
Indro Spinelli🤠CoRL@IndroSpinelli·Jun 23

The results? ⚡ 14× faster → 0.26s loop on a consumer RTX 3090 📈 +9.3pp avg success on MetaWorld and LIBERO ✅ 91.4% verdict accuracy, zero visual input Language-Action World model gives foresight without rendering and a way to interpret robot actions in natural language!

Post attachment
Indro Spinelli🤠CoRL avatar
Indro Spinelli🤠CoRL@IndroSpinelli·Jun 23

The student returns two things: a short language preview ("the gripper moves left and forward, approaching the object") and a Positive/Negative verdict. The verdict drives rejection sampling.

Post attachment
Indro Spinelli🤠CoRL avatar
Indro Spinelli🤠CoRL@IndroSpinelli·Jun 23

So we shift from "simulate-then-act" to "describe-then-act." A privileged VLM teacher caption trajectories offline describing end-effector object interactions. A small LLM student is distilled to reproduce those outputs from policy latents and the planned action chunk alone.

Post attachment
Indro Spinelli🤠CoRL avatar
Indro Spinelli🤠CoRL@IndroSpinelli·Jun 23

The catch with visual world models isn't accuracy, it's speed. Simulating future frames and evaluating them runs ~3.7s per decision on an A6000. That's far too slow for a real control loop, where you're committing to actions (or chunks) many times a second.

Post attachment
Indro Spinelli🤠CoRL avatar
Indro Spinelli🤠CoRL@IndroSpinelli·Jun 23

Many world models for robots assumes you have to render the future to know if an action will fail. Our #ECCV26 says... you don't. You can learn to verbalise future failures just from the current policy latents and selected actions🧵