Francesco Orabona
@bremen79
Dad and associate professor at @KAUST_News.
Formerly @BU_ece, @sbucompsc, @YahooResearch, @TTIC_Connect.
ML theory&practice, obsessed with history of science
446 Following 8.3K Followers
The fundamental problem with ML conference is that there is upside to doing bad work & *zero* downside. I've unearthed paper-invalidating problems with nearly every conference's Best Paper and Orals that I've touched
(two papers forthcoming on this)
1/2
Show more
I would support even stronger quotas: No more than 5 submissions per author sounds reasonable. An author is someone who has supposedly authored, at least part of a paper. Also, authors should be evaluated by their best papers, not how many they got through the process. When I see faculty candidates or researchers with 10-15 papers in one conference, it now counts as a negative signal to me.
Show more
#AAAI2027# reportedly received at least 44,000 abstract submissions.
Opinion: The publish or perish era may be over. We have reached a point where publishing alone carries almost no signal
I will give an invited talk on this result at the NExT-Game workshop at ICML tomorrow
Join me if you are interested in knowing the details of the proof, the mistakes in previous papers, and the ChatGPT proof.
Show more
I solved an open problem in optimization with ChatGPT Pro.
Consider min-max problems, where we have to minimize with respect to some variable and maximize with respect to other variables to find a saddle-point.
Show more
...too many papers...
#ICML2026#
new post on harness engineering for AI self-improvement:
It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple.
Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.
Show more
If you are at ICML, the poster (spotlight) will be at 2pm.
Come to say hi and talk with Tuo
Me: LLMs are getting so good they might replace PhD students soon.
Him: Actually, there’s an easier job in academia to automate: the PhD supervisor.
I am going to ICML, ping me if you want to chat. I am particularly interested at meeting people from frontier labs.
On Monday @natalie_collina, Ira, and I are giving a tutorial at ICML on (multi)calibration and its applications. You can find slides and an annotated bibliography on the website: as well as an interactive demo of online calibration algs. See you there!
Show more
Yesterday at the COLT business meeting, one senior core member of this community said that the COLT papers are "disconnected from reality".
The world is really changing.
Our COLT 2026 paper!
A single stepsize with high probability gives both fast and robust rates in TD learning, without using any projection.
IMO, Wei-Cheng put together a very interesting proof, with some new tricks as well.
Show more
1/11
New paper: A Single Stepsize Suffices for Unprojected Linear TD(0)
Can TD(0) without projection and knowledge of curvature achieve high-probability rates that are both robust and fast under Markovian sampling?
We show: one curvature-free stepsize + PR average suffice. 🧵
Show more
Final version of my book (with a new title)
Online Learning: A Modern Introduction Using
Convex Optimization
Especially proud of the Foreword by @NicoloCB!
It'll be printed by Cambridge University Press.
The end of 7 years of updates :)
Show more
I solved an open problem in optimization with ChatGPT Pro.
Consider min-max problems, where we have to minimize with respect to some variable and maximize with respect to other variables to find a saddle-point.
Show more
