Pierfrancesco Beneventano

This account is verified. Learn more

@PierBeneventano

Postdoc at @MIT | ex PhD student @Princeton | Thinking of how to frame theoretically agents and using agents for theory research, while brewing my espresso.

Cambridge, MA pierbeneventano.github.io Joined February 2020

2K Following 1.4K Followers 193 Posts

Latest posts 5

Pierfrancesco Beneventano @PierBeneventano Oct 18

Which LLM should answer your prompt?

A cheaper model may be enough. A more capable one may be worth the extra cost.

LLM routing makes that choice automatically, prompt by prompt.

But how does a router decide—and what happens on unfamiliar tasks?

1/9

Pierfrancesco Beneventano @PierBeneventano Oct 11

Landmark paper, very proud to share my office and coffee break with them :)

McGovern Institute Verified account @mcgovernmit Oct 9

Machine learning networks can learn without detailed instructions from a designer, according to a new study by computational neuroscientists at MIT. Those that do, may offer insight into how the brain reorganizes itself during learning.

mcgovern.mit.edu/2026/10/05/how…

View quoted post on X

Pierfrancesco Beneventano @PierBeneventano Sep 24

Got 4 papers into Neurips2026 and I'm very excited about all of them! I'm very happy that's so much better than I could have expected! Stay tuned I'm gonna write a x-piece on this research line (but after ICLR)!

Pierfrancesco Beneventano @PierBeneventano Jul 31

Major breakthrough by @alexmeterez and @alex_damian_

Alex Meterez @alexmeterez Jul 29

1/8 How well does the simplest model in optimization theory, the quadratic, actually describe LLM pretraining? Surprisingly well. We linearize a 150M model at checkpoints throughout training and the local Taylor expansion tracks the true loss for up to 10% of the training budget.
View quoted post on X

Pierfrancesco Beneventano @PierBeneventano Jul 5

Flying to Seoul to present:
arxiv.org/pdf/2604.14108 (+4 workshops) at ICML!

Let me know if you’re here and would like to chat!