Sumit Bhatia

Senior ML Scientist, Media and Data Science Research Lab, Adobe Inc.

Adjunct Prof. @IIITDelhi.

Information retrieval, Knowledge Graphs, NLP.

Horology.

[Debasis Ganguly @debforit]

Why should a simple factoid question and a complex multi-hop question receive identical treatment?

We introduce DRAG, which dynamically chooses the retriever and the generator yielding better answers and/or lower latency.

https://t.co/HtTBLKTGYg
@PayelSantra17 @sbhatia_

[Subbarao Kambhampati (కంభంపాటి సుబ��బారావు) @rao2z]

My take on AI Kumbh Mela : They went out of their way to make it possible for people from many walks of life to attend.. #IndiaAIImpactSummit2026

So going by the media accounts, all that really happened at the summit were the long security lines, lost wearables, and other more pressing first world problems like some of us not getting our millet foam at @PMOIndia dinner.

Perhaps there is another way to look at it. This is the first of the AI Summits that seemed to have allowed basically free registration to anyone who wanted to come. As far I know, the other ones in UK, Paris and Seoul were limited to mostly invited "delegates".

Opening the summit up this way meant LOTS of people showed up (as they apparently do to the other Kumbh Mela..)--estimates of registrations at 250K and daily attendance of as much as 70K. It's in a different plane all together compared to the "invited delegates only meetings", and in a different league even compared to the mega AI conferences like #NeurIPS2025.

The bigger numbers also meant longer lines, more chaos and lower signal to noise ratio for the cognoscenti.

After my panel, I met one student who brought his mother (who didn't seem particularly tech savvy) to show and tell her about AI..

I talked to a neurologist from the capital region, who showed up just to get a sense of how and whether this technology might atrophy our own cognitive skills.

A lady working in Arts and Crafts, who was trying to get a sense of how AI will affect the artists and their livelihoods.

I saw lines of women--clad in their finery--waiting for their group buses after a trip to the summit.

And I of course saw tons and tons of UG students from Indian colleges attending sessions and trying to make sense of things (however primitive some of their understanding seemed after a minute of talking to them).

Maybe some of this was orchestrated. But I would think that if AI is supposed to be such a transformative technology that would impact everyone, perhaps it is quite justified to "let everyone in"..

For that inclusiveness, I believe the organizers of this AI Kumbh Mela deserve a huge amount of credit--and our benefit of doubt on the attendant inconveniences.

(No, I have no connection with the organizers. I own my opinions.)

[Sumit Bhatia @sbhatia_]

Excited to share I didn't submit any paper to AAAI.

#AAAI #AI

[Kowndinya Renduchintala @KowndinyaR]

📢 Excited to share our latest research – “On the Effect of Instruction Tuning Loss on Generalization” – which will soon appear in the Transactions of the Association for Computational Linguistics (TACL)!!

Instruction Tuning is the very first step in the post-pretraining phase of LMs – it’s what enables today’s LMs to “follow” user instructions, whether it's summarizing a news article or giving life advice in the tone of a pirate!! 🏴‍☠️

But lurking inside almost every instruction tuning recipe is a crucial detail that has largely been overlooked –
⚠️ The conventional loss is computed only on the response tokens…

Turns out, this is suboptimal.

We propose Weighted Instruction Tuning (WIT) – a simple yet effective alternative to the conventional loss that lets you assign different weights to prompt and response tokens.

🔍 Key Findings:
👉Small Tweaks ⇒ Big Gains! – A low-to-moderate (0-0.5) prompt token weight in combination with a moderate-to-high (0.5-1) response token weight significantly boosts generalization.
👉Not only do WIT-finetuned models demonstrate consistent improvement in generalization over conventional instruction-tuned models (average relative gain of 6.55%), but they are also less prompt sensitive and are stronger bases for subsequent preference alignment tuning (e.g., DPO)!

🔧 After all, the instruction tuning loss is indeed a dial worth tuning!

✨ For a quick overview, check out my blog post at https://t.co/ZA87z94HZj
🔬For a deeper dive, read the full paper at https://t.co/HcLknxkRec
💻Try it out yourself, code available at https://t.co/xpKMi6vvDT

Finally, a huge shoutout to my amazing co-authors @anwoy_ , @sbhatia_ and @Tanmoy_Chak 🙌

#AI #ML #NLP #LLM #InstructionTuning #Research #TACL

[ISRO @isro]

GSLV-F16/NISAR

From a majestic liftoff to the flawless separation, witness the full journey.

Watch spectacular moments of NISAR launching aboard GSLV-F16 and its precise separation, captured on-board.

A milestone in global space collaboration.

#ISRO #NASA #GSLVF16 #NISAR

[Pedro Domingos @pmddomingos]

The number of papers in AI conferences has grown exponentially in the last decade. The number of good papers has stayed constant.

[Sumit Bhatia @sbhatia_]

#SIGIR2025 paper alert :)

[Debasis Ganguly @debforit]

Glad to share that "Exploring the Role of Diversity in Example Selection for In-Context Learning"
w/ @JanakKapuriya, @kaushik_manit, and @sbhatia_
has been accepted as a short paper in #SIGIR2025.

Arxiv and git link coming soon.

[Tanmoy Chakraborty @Tanmoy_Chak]

**Kindly consider sharing the post**
We are seeking opinions about the current quality of reviewing in *CL conferences. We (@emnlpmeeting PCs along with @ReviewAcl EiCs) are committed to improving the review quality. We are bringing a series of changes in the review process. Kindly consider filling out the form:
https://t.co/eQ7uKd3Cx2
@VioletNPeng

[Han @HanchungLee]

@abacaj rag isn’t “solved” because retrieval can’t be “solved”. it’s frequently a precision-recall and latency tradeoff.

that said, some of the domains are better solved than the others. metadata tags and keywords. and lots of metadata tags and keywords.

[ISRO @isro]

🎥 Relive the Liftoff! 🚀

Experience the majestic PSLV-C60 launch carrying SpaDeX and groundbreaking payloads. Enjoy breathtaking images of this milestone in India’s space journey! 🌌✨

#SpaDeX #PSLV #ISRO

📍 @DrJitendraSingh

[Pedro Domingos @pmddomingos]

Overheard at a large tech company: “We have agents for everything now. We took our templates and renamed them agents.”

[Sumit Bhatia @sbhatia_]

Can o3 compute correct GST rates? I doubt. I think that's the ultimate benchmark OpenAI and other labs should focus on.

[The Kaipullai @thekaipullai]

Google's Willow quantum computing chip has the ability to solve problems that would take supercomputers, 10 septillion years to complete

Thank God they launched before our Govts new GST rules

I think You will need every bit of its computing power to calculate the GST on popcorn

[Mark Riedl @mark_riedl]

A Nobel Prize for AI research falling under the category of Physics makes perfect sense because AI is the study of how to make silicon rocks think, and that's geology, which is a physical science, and the same root "physic" is used in both physics and physical science. I guess?

[Pedro Domingos @pmddomingos]

Boltzmann machines - Hinton’s extension of Hopfield networks - are great. The only problem is they don’t work.

[Sumit Bhatia @sbhatia_]

@fooobar Sampling bias :)

[Sumit Bhatia @sbhatia_]

People were worried about ChatGPT hallucinating. They didn't know OpenAI will be a source of unimaginable drama. Harder to align people than models, it seems.

[Pedro Domingos @pmddomingos]

When will AI reach PhD-level intelligence? Depends on the field. In fields with negative intelligence (e.g., gender studies), a pocket calculator is already far ahead.

[Sumit Bhatia @sbhatia_]

@debforit @MrParryParry @nilanjansb Congrats :)

[Jen Heemstra @jenheemstra]

No proposal or manuscript is ever perfect, but to be effective, it does eventually need to be finished.

Nothing we submit will ever be as good as it *could* be and just recognizing that can be an important step forward.

Perfection is not the goal. Finished is the goal.