Skip to main content
International Conference on Learning Representations

International Conference on Learning Representations

Search complete. 13 mentions across 13 episodes found for "International Conference on Learning Representations".

Nov 22, 2026

Mayur NaikGUEST
39:17
We have deduplicated it.
Mayur NaikGUEST
39:19
Is it now safe to use it? In our own work this year, which appeared at ICLR or Code Trek, we started out with this synthetic data set.
Mayur NaikGUEST
39:35
So again, we are looking at variable misuse detection in Python.
Mayur NaikGUEST
39:39
We started out with real programs that are written by humans.
HackerNoon AINARRATOR
45:55
Low-rank adaptation of large language models.
HackerNoon AINARRATOR
45:58
ICLR, 1, 2, 3, 2022.
HackerNoon AINARRATOR
46:01
Xiaohan Huang, Furu Wei, Lei Kui, Xingxing Zhang, and Ming Zhou.
HackerNoon AINARRATOR
46:07
Unsupervised fine-tuning for text clustering.
Stef van GriekenGUEST
19:50
I won't go into too much detail there.
Stef van GriekenGUEST
19:52
We presented it at ICLR, I think last summer, and it was one of the 12, I think, papers that were presented at that conference in the Foundation Models for Biology.
Stef van GriekenGUEST
20:05
In terms of the training procedure, so there is three steps.
Stef van GriekenGUEST
20:10
So the first step that we go through is we want to generate variants.
GustavoHOST
0:57
He went on to become a Kaggle Grand Master, serve as CTO of Solaris Therapeutics, and co-founded Olden Labs, which is a hardware robotics spin-off focused on automating animal research.
GustavoHOST
1:10
His research has been published in Neuro IPS, ICLR, and in Nature Genetics.
GustavoHOST
1:16
Today, he's the co-founder and CTO of Cosmon, building AI agents that understand engineering intent and operate CAD and simulation tools autonomously.
GustavoHOST
1:25
Welcome, Noah.
Ashwinee PandaGUEST
46:20
So this is privacy-preserving in-context learning for large language models.
Ashwinee PandaGUEST
46:26
Another paper that I'll be presenting next week at ICLR.
Ashwinee PandaGUEST
46:29
So let's say that we have some sensitive database of records, and we want to use data from this database in order to answer questions.
Ashwinee PandaGUEST
46:38
So for example, you have a document question answering question.
speaker_0HOST
0:14
Today we have Javier and Yiming for PhD students at CMU and ETH.
speaker_0HOST
0:20
And they're going to be telling us about a paper that just got accepted to ICLR on persistent pre-training poisoning of LLMs.
Javier RandoGUEST
0:31
Good.
Javier RandoGUEST
0:32
Yeah, thank you so much for the invite.
Michael AerniGUEST
30:46
And yet we seem to have sort of we seem to sort of generate text that exists verbatim on the Internet, which is something is going on.
Michael AerniGUEST
30:55
And what is going on, we call in some other recent work to be presented at ICLR this year, we call this non-adversarial reproduction.
Michael AerniGUEST
31:03
Just to make the terminology a bit clear, reproduction means an LLM or you also use it for humans generates or generates verbatim training data.
Michael AerniGUEST
31:15
That is just a reproduced text snippet is a snippet output by an LLM that exists verbatim in its training data.
Tal KachmanGUEST
24:24
You already see it.
Tal KachmanGUEST
24:25
ICLR had 62K papers submitted.
Tal KachmanGUEST
24:32
I'm now an area chair on a conference called AMAS, which is the multi-agent system conference.
Tal KachmanGUEST
24:38
It was one of my favorite conferences because it was very small and high quality.
Yacine MahdidHOST
1:18
Stuff like why planning might already be mostly solved and where the real bottleneck in robotics might be, why almost no data is bad data anymore, as long as it's labeled right, how frontier models actually are controlling these robots without training on them, whether VLA are dead now that blackboard models are actually something that works, and why robotics is just harder than language.
Yacine MahdidHOST
1:41
I'll link the Le Robo repo and their ICLR paper down below.
Yacine MahdidHOST
1:45
It's a great companion to this conversation.
Yacine MahdidHOST
1:47
Enjoy! Before we jump into this interview, let's talk about Arcee, the openweight lab sponsoring this interview.
Dhivya ChandramouleeswaranGUEST
64:18
I'm sure there'll be a revision in like November 26.
Dhivya ChandramouleeswaranGUEST
64:21
But at least for now, based on the papers that we're seeing in like ICLR and NeurIPS, what does it look like? And how I would encourage folks to read that paper and draw to their own conclusions on What does Eval look like in their own organizations or in their own personal projects based on the trends that we are seeing right now? So, yeah, things like bubble teaming, things like what's the granularity to the level we go into for observation and things like that.
Dhivya ChandramouleeswaranGUEST
64:49
Like all of that is laid out on the paper.
Dhivya ChandramouleeswaranGUEST
64:51
I think it's a fantastic paper.

3 more episodes mention International Conference on Learning Representations.

Create an account to see the whole feed, search across every transcript, and follow the entities you care about.

We value your privacy

We use cookies to understand how you use our platform and to improve your experience. Click “Accept All” to consent, or “Decline non-essential” to opt out of non-essential cookies. Read our Privacy Policy.