Aditya Gopalan

v2.0: Casually eating bread is being naan challahnt

@CsabaSzepesvari @vishalmisra Adding to @CsabaSzepesvari 's remark, in the HMM expts section, the entropy computed as HBayes(t), after eq 10, is also not averaged across the posterior dist of the HMM params (E,T) to form the actual posterior predictive distribution.

@vishalmisra The alpha values are the posterior probability given the prior over the state and the model E,T. But in the setting you describe, theta=(E,T) is random, from an (unspecified) prior. The actual posterior for one thing would involve this prior. 2/2

New work: Do transformers actually do Bayesian inference? We built “Bayesian wind tunnels” where the true posterior is known exactly. Result: transformers track Bayes with 10⁻³-bit precision.
And we now know why. I: arxiv.org/abs/2512.22471 II: arxiv.org/abs/2512.22473
🧵

What did the Gujju order amidst hunger pangs while bathing? . . . . . Shawarma.

@farairesearch Thank you @farairesearch A preprint of this work is now available: arxiv.org/abs/2510.20413

"DPO can give you a policy that is worse than what you started with" @today_itself reveals how the theoretical backing behind one of the most ubiquitous alignment methods breaks down for real-life LLMs, causing unpredictable alignment failures. He then shows how to fix it.

May your AI train on good slop

How did the dosa NOT want to be seen sitting? Idly.

@DimitrisPapail “Studio glibly”

I asked ChatGPT to generate the image of the first page of its favorite ML paper. This is what it produced :)

@HaqueIshfaq Same question here

What are we supposed to put under "PARTICULARS OF LOCAL CONTACT" in Form14a for Singapore visa to attend ICLR conference? It seems, it is mandatory to put something there? ICLR website says, they can't provide a local contact 🤷‍♂️ #ICLR2025 #Singapore

@rao2z @iclr_conf Welcome to visit IISc if you plan to be in BLR. You can enjoy badami and other local varieties ... err ... interact with AI researchers on campus.

I am planning to go to @iclr_conf #ICLR2025 in Singapore and might stop by in India after the conference (April end--May beg) to enjoy mangoes err.. interact with AI researchers in India. Please DM if you have suggestions..

Heard from an academic: “They work on SOTA models, I work on chhota models.”

The Walmart Centre for Tech Excellence at @IIScCSA proudly presents the Reinforcement Learning Workshop 2025. Join us for a live stream on YouTube: youtube.com/watch?v=wnhgJE…. For the full schedule and more details, visit events.csa.iisc.ac.in/rlworkshop2025/. Don't miss out!

Omg. ((2+0+2+5)×(2-0-2-5))²=2025

@imrahulmaddy I even predict the next word in my head …

what? people don't hear words in their head as they read?

academics are so cute. look how they write little papers for each other to read. they even format it all nice. aww

@MountainOfMoon Conferences should just make this phrase a radio button. It is literally the bulk of messages.

This is called getting friendzoned by a reviewer

A couple of years ago, I had taken a low resolution image of this. Here.

@iissarayu @ICICIBank_Care Wonder if consulates in India have giant warehouses solely dedicated to the storage of printed, rubber stamped and signed 6-month bank statements …

Okay - @ICICIBank_Care is a … mess up. I’ve been waiting for 1.5 hours for a bank statement (prints needed in 2024 only for visa applications, another story)

v2.0: Casually eating bread is being naan challahnt

Casually eating bread is being naan chalant

Considering asking ECE-IISc to adopt this as its motto …

@ccanonne_ Did you look into the mirror to see your descent?