Shubhendu Trivedi
@shubhendu.bsky.social
Interests on bsky: ML research, applied math, and general mathematical and engineering miscellany. Also: Uncertainty, symmetry in ML, reliable deployment; applications in LLMs, computational chemistry/physics, and healthcare.
https://shubhendu-trivedi.org
https://shubhendu-trivedi.org
Reposted by Shubhendu Trivedi
New technical blog post on the Learning Theory Alliance blog, by Rohith Kuditipudi, on watermarking language models.
A super interesting problem where theory is essential! Please check it out.
www.let-all.com/blog/2025/11...
A super interesting problem where theory is essential! Please check it out.
www.let-all.com/blog/2025/11...
November 26, 2025 at 3:53 PM
New technical blog post on the Learning Theory Alliance blog, by Rohith Kuditipudi, on watermarking language models.
A super interesting problem where theory is essential! Please check it out.
www.let-all.com/blog/2025/11...
A super interesting problem where theory is essential! Please check it out.
www.let-all.com/blog/2025/11...
Reposted by Shubhendu Trivedi
How do brain areas control each other? 🧠🎛️
✨In our NeurIPS 2025 Spotlight paper, we introduce a data-driven framework to answer this question using deep learning, nonlinear control, and differential geometry.🧵⬇️
✨In our NeurIPS 2025 Spotlight paper, we introduce a data-driven framework to answer this question using deep learning, nonlinear control, and differential geometry.🧵⬇️
November 26, 2025 at 7:32 PM
How do brain areas control each other? 🧠🎛️
✨In our NeurIPS 2025 Spotlight paper, we introduce a data-driven framework to answer this question using deep learning, nonlinear control, and differential geometry.🧵⬇️
✨In our NeurIPS 2025 Spotlight paper, we introduce a data-driven framework to answer this question using deep learning, nonlinear control, and differential geometry.🧵⬇️
Just engaged with 20+ "we thank the reviewer" things for the past 3-4 hours (ICLR, ARR, AISTATS) and someone mentioned thanksgiving. I forgot.
November 26, 2025 at 9:10 PM
Just engaged with 20+ "we thank the reviewer" things for the past 3-4 hours (ICLR, ARR, AISTATS) and someone mentioned thanksgiving. I forgot.
Let us say you have some partial knowledge about a quantity of interest before any data is analyzed. Knowledge is specified by a set of lower and upper probabilities assigned to hypotheses about said quantity, and is time-indexed. Can you any-time validity with improved efficiency?
Regularized e-processes: anytime valid inference with knowledge-based efficiency gains
Classical statistical methods have theoretical justification when the sample size is predetermined. In applications, however, it's often the case that sample sizes are data-dependent rather than prede...
arxiv.org
November 26, 2025 at 8:06 PM
Let us say you have some partial knowledge about a quantity of interest before any data is analyzed. Knowledge is specified by a set of lower and upper probabilities assigned to hypotheses about said quantity, and is time-indexed. Can you any-time validity with improved efficiency?
A more formal paper on a theory of value of information www.econ2.jhu.edu/conferences/...
www.econ2.jhu.edu
November 25, 2025 at 9:43 PM
A more formal paper on a theory of value of information www.econ2.jhu.edu/conferences/...
Reposted by Shubhendu Trivedi
🤔💭What even is reasoning? It's time to answer the hard questions!
We built the first unified taxonomy of 28 cognitive elements underlying reasoning
Spoiler—LLMs commonly employ sequential reasoning, rarely self-awareness, and often fail to use correct reasoning structures🧠
We built the first unified taxonomy of 28 cognitive elements underlying reasoning
Spoiler—LLMs commonly employ sequential reasoning, rarely self-awareness, and often fail to use correct reasoning structures🧠
November 25, 2025 at 6:26 PM
🤔💭What even is reasoning? It's time to answer the hard questions!
We built the first unified taxonomy of 28 cognitive elements underlying reasoning
Spoiler—LLMs commonly employ sequential reasoning, rarely self-awareness, and often fail to use correct reasoning structures🧠
We built the first unified taxonomy of 28 cognitive elements underlying reasoning
Spoiler—LLMs commonly employ sequential reasoning, rarely self-awareness, and often fail to use correct reasoning structures🧠
I won't be able to make it for NeurIPS (will be a bit of a period of flux). But in case you happen to be there, please try to stop by and chat with Marco. This will also be a spotlight.
Our recent paper on universality classes of equivariant networks arxiv.org/abs/2506.02293 enjoyed being part of it (and more to come!).
On Universality Classes of Equivariant Networks
Equivariant neural networks provide a principled framework for incorporating symmetry into learning architectures and have been extensively analyzed through the lens of their separation power, that is...
arxiv.org
November 25, 2025 at 7:31 PM
I won't be able to make it for NeurIPS (will be a bit of a period of flux). But in case you happen to be there, please try to stop by and chat with Marco. This will also be a spotlight.
I know this is just a SM post. But one could argue that Finance (esp. the macro-structural tradition) is the more "adult" field in terms of its brutal enforcement of system-level realism than others listed. You misread the system, and you disappear. There's no safe "academic" abstraction layer. [+]
Many modern technologists and computer scientists are very cogent systems thinker because any sufficiently advanced technology *is people*, all technical systems are socio-technical. The capital T Tech dipshits talking about “alpha” are mainly cosplaying finance people, regardless of their training
This is absolutely the case. Tech Guys of previous generations, to the extent that that category can be read back into the 19th and 20th centuries, were shockingly well-read by modern standards
November 25, 2025 at 4:31 PM
I know this is just a SM post. But one could argue that Finance (esp. the macro-structural tradition) is the more "adult" field in terms of its brutal enforcement of system-level realism than others listed. You misread the system, and you disappear. There's no safe "academic" abstraction layer. [+]
Led me to a very informative wikipedia entry en.wikipedia.org/wiki/Opaque_...
November 25, 2025 at 8:20 AM
Led me to a very informative wikipedia entry en.wikipedia.org/wiki/Opaque_...
I showed this to my mother on Zoom just now and before I got an opinion on how it looked, it became about "why do I always need to collect trash." 😅 But before I could explain that I had no intention to get it, my father started to lament that she threw away all the old (even antique) rotary phones.
FYI: In trying to look up the phone model on the top post, I learnt that oldphoneworks (a reliable source for refurbished phones) no longer ships to the US. Also found this handsome phone while looking up the above (and looking up for pre-1920 non-brass examples).
Model 102 Deskset - Oldphoneworks
Western Electric Model 102 "B" Mount with E1 Handset.. Originally manufacturered in the late 1920's. We have installed an antisidetone network in the the base to ensure optimal sound quality on modern...
oldphoneworks.com
November 25, 2025 at 5:36 AM
I showed this to my mother on Zoom just now and before I got an opinion on how it looked, it became about "why do I always need to collect trash." 😅 But before I could explain that I had no intention to get it, my father started to lament that she threw away all the old (even antique) rotary phones.
Just recently got an invite for a workshop in Sarajevo, and then another one immediately after to teach ML at a winter complex systems school in Petnica (about an hour from Belgrade). Of course, it's just at a time when I can't travel/leave the U.S. :-/
November 24, 2025 at 2:24 AM
Just recently got an invite for a workshop in Sarajevo, and then another one immediately after to teach ML at a winter complex systems school in Petnica (about an hour from Belgrade). Of course, it's just at a time when I can't travel/leave the U.S. :-/
These can still be obtained. Although it can be a challenge to get nicely refurbished ones (unless they are of brass).
As one of the first table top phones, these were certainly an identity purchase.
These marked a shift of communication moving from the public to a more private interaction before eventually going public again.
This phone served more than a way to talk. It served as a social and practical anchor.
These marked a shift of communication moving from the public to a more private interaction before eventually going public again.
This phone served more than a way to talk. It served as a social and practical anchor.
November 23, 2025 at 7:22 PM
These can still be obtained. Although it can be a challenge to get nicely refurbished ones (unless they are of brass).
Conferences trying hard to play with semantics.
November 23, 2025 at 3:49 AM
Conferences trying hard to play with semantics.
Regret not running a sort of inverse LEAP (dated two years out), despite very strong conviction. Closed it out in one month.
November 22, 2025 at 9:04 PM
Regret not running a sort of inverse LEAP (dated two years out), despite very strong conviction. Closed it out in one month.
Merely the pattern match from the post reminds of a book from a few years ago. I liked the idea enough to pull all stops to get a copy from the publisher almost 1.5 years before it was published. This reminds me: 1) I should read it again, 2) Should get a prettier copy.
November 22, 2025 at 7:02 PM
Merely the pattern match from the post reminds of a book from a few years ago. I liked the idea enough to pull all stops to get a copy from the publisher almost 1.5 years before it was published. This reminds me: 1) I should read it again, 2) Should get a prettier copy.
हक़ अच्छा.. हक़ अच्छा.. पर उस के लिए कोई और मरे तो और अच्छा
तुम भी कोई मंसूर हो जो सूली पे चढ़ो.. ख़ामोश रहो ~ इब्न-ए-इंशा
तुम भी कोई मंसूर हो जो सूली पे चढ़ो.. ख़ामोश रहो ~ इब्न-ए-इंशा
November 20, 2025 at 11:28 PM
हक़ अच्छा.. हक़ अच्छा.. पर उस के लिए कोई और मरे तो और अच्छा
तुम भी कोई मंसूर हो जो सूली पे चढ़ो.. ख़ामोश रहो ~ इब्न-ए-इंशा
तुम भी कोई मंसूर हो जो सूली पे चढ़ो.. ख़ामोश रहो ~ इब्न-ए-इंशा
Great work by Eugène: arxiv.org/abs/2511.15146
Beyond Uncertainty Sets: Leveraging Optimal Transport to Extend Conformal Predictive Distribution to Multivariate Settings
Conformal prediction (CP) constructs uncertainty sets for model outputs with finite-sample coverage guarantees. A candidate output is included in the prediction set if its non-conformity score is not ...
arxiv.org
November 20, 2025 at 5:13 AM
Great work by Eugène: arxiv.org/abs/2511.15146
What people are actually trying to say: Natural data distributions live near a low-dimensional tame geometric object. Neural networks can represent functions whose images are also tame. PS: This implies learning is possible because both sides reside in the same o-minimal universe.
Find someone who loves you more than ML people love saying manifold
November 20, 2025 at 2:46 AM
What people are actually trying to say: Natural data distributions live near a low-dimensional tame geometric object. Neural networks can represent functions whose images are also tame. PS: This implies learning is possible because both sides reside in the same o-minimal universe.
Bienvenidos Inteligencia Artificial Al Servicio De Las Personas.
(Doesn't mean anything, just a phrase stuck in my head since morning).
(Doesn't mean anything, just a phrase stuck in my head since morning).
November 19, 2025 at 2:39 AM
Bienvenidos Inteligencia Artificial Al Servicio De Las Personas.
(Doesn't mean anything, just a phrase stuck in my head since morning).
(Doesn't mean anything, just a phrase stuck in my head since morning).
Reposted by Shubhendu Trivedi
Alternate between AlphaFold3 "hallucinations" and fixed-backbone sequence design to generate or optimize proteins.
www.biorxiv.org/content/10.1...
www.biorxiv.org/content/10.1...
November 18, 2025 at 10:39 PM
Alternate between AlphaFold3 "hallucinations" and fixed-backbone sequence design to generate or optimize proteins.
www.biorxiv.org/content/10.1...
www.biorxiv.org/content/10.1...
An important reason to fix reviewing (solution is simple, by the way: phase out double blind) is to reduce the whole "If it were not for rectangular-headed reviewers, we were winning the Nobel Prize. Scholarship is dead, etc." shtick. Tired of this every cycle.
November 18, 2025 at 11:30 AM
An important reason to fix reviewing (solution is simple, by the way: phase out double blind) is to reduce the whole "If it were not for rectangular-headed reviewers, we were winning the Nobel Prize. Scholarship is dead, etc." shtick. Tired of this every cycle.
Reposted by Shubhendu Trivedi
❤️ Love this ❤️
@mathforge.org has done what I’ve never been organised enough to do myself, and catalogued every #geometrypuzzle I’ve ever posted on social media. It looks amazing!
@mathforge.org has done what I’ve never been organised enough to do myself, and catalogued every #geometrypuzzle I’ve ever posted on social media. It looks amazing!
November 17, 2025 at 6:33 AM
❤️ Love this ❤️
@mathforge.org has done what I’ve never been organised enough to do myself, and catalogued every #geometrypuzzle I’ve ever posted on social media. It looks amazing!
@mathforge.org has done what I’ve never been organised enough to do myself, and catalogued every #geometrypuzzle I’ve ever posted on social media. It looks amazing!
Index space estimation (both single-/multi-index) used to be a niche area of semi-parametric statistics--had strong intellectual lineage, but was still isolated, and disappeared in the SVM era. It seems to keep resurfacing in ML now, and I am not quite able to understand why.
November 17, 2025 at 2:02 AM
Index space estimation (both single-/multi-index) used to be a niche area of semi-parametric statistics--had strong intellectual lineage, but was still isolated, and disappeared in the SVM era. It seems to keep resurfacing in ML now, and I am not quite able to understand why.
The solution is to phase out double blind reviews, make reviewing compulsory for authors. f1000 shows how this should be done. Nature also gave you the option to reveal your identity after the process was complete (I see that ARR has been experimenting with it last two cycles).
All the slop review drama unfolding at ICLR is painful to watch. But having so many eyes on it puts us in a better point to find solutions. This would have taken longer to be exposed on a closed platform where we may have had a few screenshots and a few author stories at best.
November 16, 2025 at 10:06 PM
The solution is to phase out double blind reviews, make reviewing compulsory for authors. f1000 shows how this should be done. Nature also gave you the option to reveal your identity after the process was complete (I see that ARR has been experimenting with it last two cycles).
Something reminded me of books published under the "Writings from an unbound Europe" series. Some years ago I had learnt it had gone out of print (10 y prior), and that explained to me why every other book randomly seemed to be unavailable. So I decided to track down and get all of the volumes, [+]
November 16, 2025 at 6:22 PM
Something reminded me of books published under the "Writings from an unbound Europe" series. Some years ago I had learnt it had gone out of print (10 y prior), and that explained to me why every other book randomly seemed to be unavailable. So I decided to track down and get all of the volumes, [+]