The FORC 2026 call for papers is out! responsiblecomputing.org/forc-2026-ca... Two reviewing cycles with two deadlines: Nov 11 and Feb 17. If you haven't been, FORC is a great venue for theoretical work in "responsible AI" --- fairness, privacy, social choice, CS&Law, explainability, etc.
It's time for the Theory Jobs spreadsheet for 2025! The theory CS community crowdsources who accepted which jobs every year.
Links in next post 1/2
W2
New post: "Apple Machine Learning Research at #ICLR 2025" - highlighting a selection of the many Apple #ML research papers to be presented at @iclr-conf.bsky.social this week: machinelearning.apple.com/research/icl...
Check out our Apple research work on scaling laws for native multimodal models! Combined with mixtures of experts, native models develop both specialized and multimodal representations! Lots of rich findings and opportunists for follow up research!
Incidentally, cosh also turns out to be an excellent potential function when analyzing more general balls and bins processes, including 2-choice, and processes on graphs.
www.microsoft.com/en-us/resear...
That’s fair. Ok the mac, I use a default “scratchpad” note on the Notes app for such purposes. I like the simple interface and seamless sync. I can imagine it being similarly easy on the phone if one optimized for this use case.
The existence of this option unfortunately also means that the market for your dream device is likely too small.
If you usually carry a phone with you, that should be able to serve this purpose fairly well (and this use should not use up too much of your battery).
I think there are no easy solutions here, except for some kind of rotation. US visa is hard for many, but there is also a large group of students with single-entry US visas that are in the US but cannot easily leave and come back.
📌
Longer post No conference deadline. Likely a strong desire for the post itself to be very well-written. The odds are not in our favor here :)
Overall, evidence for the claim “And motor vehicle accidents began to climb sharply” is missing in the article. Perhaps there was more that was chopped off during editing. But as published, the article seems to make a strong assertion without evidence backing it. (4/4)
The claim “Statewide, crashes on roads patrolled by the State Police jumped 18 percent between January and March.” reeks of cherry picking. Different set of roads from the last stat. The months compared make no sense from the point of view of the argument being made. (3/4)
The claim "The number of accidents increased, year over year, for six of the next seven months." suggests that there is data somewhere on accidents beyond August. But without numbers on the increase, this claim is meaningless. Also, we want accidents/mile driven here. (2/4)
This was a very interesting article, but I wish the argument for impact was better. There is one quoted figure of 27% increase (relative to what?) which I did not see in the linked document. The graphs on page 61 there seem to show no noticeable change. Perhaps this info is somewhere else. (1/4)
I think some prescribed limits on how much time and effort authors/reviewers are expected to spend during this phase would be useful. I'd say it's fine if authors want to spend a lot more time during rebuttal, rather than resubmit. But I would much prefer that be the exception, not the expectation.
There are also papers that are already good, but get better during the process, which is a good outcome. The downside is the additional overhead for papers that are ultimately rejected, and that for papers that would be accepted anyway.
I think the greatest benefit of the rebuttal cycle is that papers that would otherwise be rejected and resubmitted and accepted to the next conference are removed from the system sooner. This reduces load on the system as a whole and thus improves outcomes overall.
…the question you are probably asking is “If we did not have rebuttals, and got to a new equilibrium, how would that compare.” This is more complicated as both authors and reviewers would behave differently in that world.
I have seen rebuttals both improve papers and lead to better decisions, given the current system. And in the equilibrium we are in, I think this is fairly common. But…
Repost if BlueSky is now your primary social media site.
I have always thought that the green flash is an in joke amongst island travel guide writers. Can I br sure you are not writing one?
“The starter pack of MLers who are not in any starter pack”
There is variance from conference to conference admittedly, and also across sub-areas. I think 5% is quite an outlier, and is perhaps more a sign of paper matching failure.
While incentives would help, I will point out that in slow-growing communities, reviewing does largely work fine. The problem with ML is the fast growth and the matching problem. In small sub-communities in ICLR/ICML/Neurips, more than half the reviews are usually reasonable, in my experience.
Perhaps there is also some additional impact of how agents behave. The network effect manifests itself by agents in the network engaging. If agents in B were to (temporarily) strategically increase the quality and quantity of their engagement, they could make a big difference in outcomes.
I think these are different questions as an author vs. as an attendee. For attendee optimization, how about trying a combination of (A-) and (B++): k=1; make talks 6-8 minutes and have long (mandatory for authors) poster sessions for deeper discussions. Add longer online talks.
Some of our algorithms have run times. Some have walk times. Others have what can only be described as crawl times.
Other conferences: At least one author of the paper is required to register. If you don’t register by tomorrow, we will remove your paper from the proceedings.
Neurips: Yo, we noticed no one from your paper registered. That rocks as we were short on poster space anyway.