@sidgairo18 This is exactly the mess that TMLR solved. There are no scores in round-1 of reviews. There is only yes/no decision after rebuttals. There are enough reasons why every system needs to converge to this.
English
Anirbit
172 posts
Anirbit
@anirbit_maths
Lecturer in ML, The University of Manchester Action Editor @ TMLR Associate Editor @ ACM-TOPML
Katılım Ocak 2025
107 Takip Edilen52 Takipçiler
Food for thought - 🤔
I've been thinking about this long and hard - having been reviewing for popular ML / CV conferences (ICML, ICLR, NeurIPS, CVPR, ICCV, ECCV) - with the community submitting papers across these, it only makes sense to have a uniform reviewer form, guidelines, rules and format across these conferences.
Personally I have a real hard time calibrating my scale from 1-10 (ICLR) to 1-6 for CVPR, then we comes ICML which also has 1-6 but 3,4 are weak reject/accept instead of 3,4 as borderline reject/accept (for CVPR).
This only gets trickier and worse when you add ICCV, ECCV, NeurIPS into the mix. Then, you add NLP related conferences and Robotics ones, to make the entire system more and more confusing - with uncalibrated reviewer scores coming - which may or may not truly reflect the reviewer's intentions.
Happy to hear the thoughts of others.
cc: @icmlconf @CVPR @NeurIPSConf @iclr_conf @ICCVConference @eccvconf
English
India's #Sarvam LLMs are among the best in class LLMs in these size categories. Both are reasoning models implementing Grouped Query Attention (GQA) and Multi-Head Latent Attention (MLA).
Pratyush Kumar@pratykumar
📢 Open-sourcing the Sarvam 30B and 105B models! Trained from scratch with all data, model research and inference optimisation done in-house, these models punch above their weight in most global benchmarks plus excel in Indian languages. Get the weights at Hugging Face and AIKosh. Thanks to the good folks at SGLang for day 0 support, vLLM support coming soon. Links, benchmark scores, examples, and more in our blog - sarvam.ai/blogs/sarvam-3…
English
@akshayrangamani Hello Akshay 😁 For a start, gradient-flow is a PDE! Way too much of modern AI is hinged on understanding gradient-flows. Then SDEs, underlie so much of noisy-S/GD, which we understand by their density evolution by FPS PDE.
English
@anirbit_maths Isn't the post saying PDEs aren't really represented in modern AI?
What kind of PDE concepts are you thinking of?
English
Nice set of books, but not sure if many AI experts know **all** these topics.
dr. jack morris@jxmnop
it always disappointed me that such a small subset of mathematical ideas matter for AI i miss doing real math
English
I mentioned this "test of AGI" in so many conversations 🙄 Only when @demishassabis says it, people start to talk about it 🙂
Rohan Paul@rohanpaul_ai
Demis Hassabis’s “Einstein test” for defining AGI: Train a model on all human knowledge but cut it off at 1911, then see if it can independently discover general relativity (as Einstein did by 1915); if yes, it’s AGI.
English
Live translation of lectures by #SarvamAI was a highlight at the #IndiaAISummit2026 🔥
RapperPandit@RapperPandit
🚨My God ! I was not understanding why this Video is Viral , the Prof is just giving a Lecture normally shifting from one language to other. Then I come across the post by @thebetterindia . And I came to know that it was a Realtime translation.
English
@thegautamkamath My anecdotal evidence is that rebuttals in general (for majority?) do no good to the authors. Is there statistics on how many papers crossed the acceptance threshold after rebuttals?
English
Suppose one of NeurIPS/ICML/ICLR decided to do away with all rebuttals. Acceptances/rejections would be decided by the reviewers and the ACs, without input from the authors beyond the submissions. Which would you, as an author and a reviewer jointly, prefer?
English
My post-doc mentor gets the COPSS award!
Incredible!
Congratulations to
@weijie444 !
My students are still building projects on top of Weijie's works!
community.amstat.org/copss/awards/p…
English
#MyXAnniversary 😎
It started with posting about our paper on provable training of nets of any size 😁
English
@kamalikac Is there a pathway from mechanistic interpretability to task specific neural architecture search? I would love to know your views on this 🙂
English
I am looking for more topics for blog-posts; please DM your suggestions! Topics can be anything related to AI, privacy/security/safety, generalization, LLMs, career advice. A reminder that I cannot blog about my employers or anything specific/sensitive to them.
English
@andrewgwils I thought overparametrization is a silver bullet. Then LLMs happened where almost always, it seems #training-tokens >>> #parameters ?
English
Have you ever reversed your position on a strongly held technical belief? What was the belief and what convinced you to change your mind?
English
Seems our work on using Villani functions to prove neural training is now among the most-read recent papers in the IMA journal, II! 💥
This is the first (only?) truly “beyond-NTK” proof. Do check our related recent work, arxiv.org/abs/2503.10428…
English
This is the key difference between in-domain and out-of-domain generalization, and we still have not truly solved out-of-domain generalization. It just turns out you can build world changing technology by throwing so much data at things that the entire universe is in-domain.
Niels Rogge@NielsRogge
One of the best visual explanations I've ever seen for why scaling Transformers works, but is suboptimal, as it's just brute-forcing things, by @YesThisIsLion (co-author of the Transformer) on @MLStreetTalk "In the (rejected) paper "Intelligent Matrix Exponentiation", they show the decision boundary of a classic MLP with a ReLu/Tanh activation function on the classic Spiral dataset." "You can see they both technically solve it with great scores on the test set. Next, they show the decision boundary of the "M-layer" they propose in the paper. And it represents the spiral ... as a spiral!" "Shouldn't we? If the data is a spiral... shouldn't we represent it as a spiral?" "If you look back at the decision boundaries of the MLP, it's clear that you just have these tiny, piecewise separations without learning the concept of a spiral. That's what I mean!" "If you train these things enough, it can fit the spiral and get a high accuracy. But there's no indication that the MLP actually understands a spiral. When you represent it as a spiral, it extrapolates correctly, cause the spiral just keeps going out."
English
This was the slide where I outlined the 2 key questions which I think are foundational to progress with neural operators & #AI4Science . ACM IKDD #CODS2025, gave a platform for such discussions between new academics and subject stalwarts in the audience 💥
Anirbit@anirbit_maths
Gave my "new faculty highlight" talk at the ACM IKDD #CODS 2025 - where I outlined a vision for neural operator research - and reviewed our 2 #TMLR papers from 2024, in the theme.
English
@deepcohen As much as I agree with the idea of giving "falsifiable predictions", I feel it would be near-impossible to publish such papers. A lot of theory is cool to do but not done because of this official requirement. I hope to be wrong about this!
English
So, we should focus on theories that can reliably predict “the small things” about deep learning, and gradually broaden the scope of what we can predict, until we have theory that can reliably predict “the big things” about deep learning too.
English
The goal of deep learning theory/science is to guide practice. But most practical questions are >1 paper away from being legitimately answered by theory. How, then, can we make progress, without access to the ideal reward signal of “does this theory give us a SOTA algorithm?” …
English
@PreetumNakkiran Ofcourse one can't rule out that tomorrow a training guarantee might emerge that critically leverages some subnet property. If that happens, that could be a serious bolstering of the view of mechanistic interpretability.
English
@PreetumNakkiran "understand" means opposite things to these 2 communities. As someone wanting provable training, I am far less concerned about the complexity of my proof. Mechanistics are hoping the otherway, that key features of big models somehow exist in simpler subnets.
English
“Theory of deep learning” went through similar discussions about its goals & purpose some ~5yrs ago. Someone should write about the relations between mech-interp & theory: two communities w/ fundamentally similar motivations (“understand neural nets”), but very different methods.
David Bau@davidbau
At the #Neurips2025 mechanistic interpretability workshop I gave a brief talk about Venetian glassmaking, since I think we face a similar moment in AI research today. Here is a blog post summarizing the talk: davidbau.com/archives/2025/…
English
@gowthami_s David Pfau's results are from years ago. It was clear right then that AI-for-Science works 🙂
English
