Sumit Kumar

Research student @iiit_hyderabad @precogatiiith. I like KGs.

Representation & Relational Learning | Foundation Models (for structured data) | Graph ML

[Sumit Kumar @sumitkk01010]

Current KGFMs ignore the rich textual signals, mostly present in KGs, and focus only on structure. We show that relation text similarity acts as a transferable invariance. Enriching it with LLMs unlocks stronger generalization 🚀
Check the thread below for more on our new work!

[Arvindh Arun @arvindh__a]

Does text help KG Foundation Models generalize better? 🤔

Yes (and no)! ☯️

Bootstrapped by LLMs improving KG relation labels, we show that textual similarity between relations can act as an invariance - helping generalization across datasets! 🧵👇

[Ponnurangam Kumaraguru “PK” @ponguru]

🤝🤝 Connections in #Bloomington, #Indiana, please welcome @sumitkk01010 visiting the @IndianaUniv.

👋🏾If you are around, interested in talking about #LLMs #RelationalLearning #KowledgeGraphs or generally catching up w/ @precogatiiith students, feel free to say hello. #ProfGiri

[Sumit Kumar @sumitkk01010]

We’ll be presenting SEMMA at #EMNLP2025 (Suzhou, China)!

📍Session 13, Hall C — November 7 (10:30–12:00)

If you’re attending @emnlpmeeting and interested in #Ontologies, #KnowledgeGraphs , or #FoundationModels, do drop by and check it out!

[Ponnurangam Kumaraguru “PK” @ponguru]

✈️ Headed to @emnlpmeeting #EMNLP2025 (Suzhou, Nov 4–9)? Meet @sumitkk01010 Swapnanil @arvindh__a & Omkar Gurjar @precogatiiith presenting exciting work on KGs, VLMs & LLMs! Catch them at their 🗓️ posters on Nov 7! 📜 Paper links in comments. #PrecogsRock #ProfGiri

[Sumit Kumar @sumitkk01010]

@cneuralnetwork What rankings are these?
WF is tomorrow.

[Sumit Kumar @sumitkk01010]

SEMMA is now accepted at #EMNLP2025 Main! We worked very hard on this paper, and I had a great time collaborating with @arvindh__a @BoXiongs @MojtabaNayyeri @tetraduzione @ponguru and Steffan. Feels great to have my first co-author paper get accepted at such a great venue!

[Arvindh Arun @arvindh__a]

SEMMA (transliteration of செம்ம - meaning awesome), my first PhD work, is accepted to #EMNLP2025 Main!

I also found out today that SEMMA has the (tied) highest average reviewer score in this ARR cycle 💪

📜: https://t.co/mhb6ecROz9

[Sumit Kumar @sumitkk01010]

@michael_galkin @arvindh__a @tetraduzione @BoXiongs Additionally, SEMMA’s textual component can be integrated with KG-ICL 🙂, considered a more expressive structural baseline that leverages few-shot prompt graphs specific to each query.

[Sumit Kumar @sumitkk01010]

@michael_galkin @arvindh__a @tetraduzione @BoXiongs Thank you for the pointers!
KG-ICL is strong but needs few-shot examples per relation at test time, whereas SEMMA is truly zero-shot (Sec 5.4, App F). Computationally, KG-ICL’s preprocessing scales with # prompt graphs and # relations; SEMMA is linear w.r.t. relations (App D).

[Ponnurangam Kumaraguru “PK” @ponguru]

📢Tejasvi's #MSThesisDefense "Personal Narratives in Online Political Discussions"

Thesis📜 https://t.co/aWQDbnYmzr

Papers📜
(0) https://t.co/B0gcRQE3ZM
(1) https://t.co/imdxkNkTPf

🕐3:30PM IST, 19th May

🗾 Flyer

#OpenToAll anybody can attend the #Defense

#ProfGiri

[Ponnurangam Kumaraguru “PK” @ponguru]

Paper #1
Website [Paper, Code]: https://t.co/TLbg4pcZOe
https://t.co/SHrgyWpHu5

X 🧵 https://t.co/1Fj28atxBl W/ @ShashwatGoel7 @arvindh__a @VarshitaKolipa1 @akshitwt @DebanganM10375 @sumitkk01010

[Ponnurangam Kumaraguru “PK” @ponguru]

Interested in #Interpretability #Explainability #Unlearning & more? Join 🧑🏽‍🏫 👨🏽‍🏫 CS7.405: Responsible & Safe AI Systems, course project (list 👇🏽) poster presentation at KRB Exhibition Hall, IIIT Hyderabad. #OpenToAll #ProfGiri 🙏🏽 #CFBR.

Course materials: https://t.co/qEJbicNLeo

[Ponnurangam Kumaraguru “PK” @ponguru]

⚽️🏀 An action-packed #Telangana State Inter-Engineering College Tournament 2025 held @iiit_hyderabad. We had matches in #Basketball (Men & Women) & #Football (Men)! Congrats to all the winners & participants! 👌 organization by the PEC & Team. #SportsatIIITH #ProfGiri 🏅🏅

[Ponnurangam Kumaraguru “PK” @ponguru]

🎉🎊📢 Can LMs not just generate solutions but also falsify incorrect ones? Falsification is key to reliable reasoning. Top models (o3-mini, R1) find counterexamples for <9% of wrong solutions in our new benchmark - REFUTE! Oral @iclr_conf Self-Improving workshop (top 6!) 🧵👇

[Ponnurangam Kumaraguru “PK” @ponguru]

Interested in #AppliedML 🤖 #Graphs #AppliedNLP #Multimodality 🎭 #Interpretability 🔍 #Fairness & #Bias ⚖️ ?

🚀 Experience #Research in the Air & 💦 Water @iiit_hyderabad Our 👌🏽 students are showcasing exciting posters at 📍 H104 today & tomorrow. #OpenToAll #ProfGiri

[Ponnurangam Kumaraguru “PK” @ponguru]

📣 #Newpaper🚨 Our work "Great Models Think Alike & this Undermines AI Oversight"
📜 https://t.co/99duwloNUK
🌐 https://t.co/e0AFyTN3EH
📂 https://t.co/fR9xVz29Nz

w/ @ShashwatGoel7 @JoschkaStrueber @AmandaIlze @karunakc_ @douwekiela @MatthiasBethge @AmyPrb @jonasgeiping
🧵👇

[Shiven Sinha @shiven_sinha]

AI can generate correct-seeming hypotheses (and papers!). Brandolini's law states BS is harder to refute than generate. Can LMs falsify incorrect solutions? o3-mini (high) scores just 9% on our new benchmark REFUTE. Verification is not necessarily easier than generation 🧵

[Sumit Kumar @sumitkk01010]

The first research project of which I was part is out! Absolutely thrilled to have been part of this project. This has been such a huge learning experience!
Unlearning memories ftw 🏁
Check out the thread below for more details.

[Akshit @akshitwt]

How do we remove the effect of incorrect training data on a trained model? Is retraining from scratch the best we can do?
We show no! Cognac 🍷 is 8x faster than retraining from scratch, and works even when you discover as little as 5% of such data.
How? 🧵👇

[Ponnurangam Kumaraguru “PK” @ponguru]

Interested in reading 📖 (open to all), discussing 🗣️🗣️ papers? 🚀 Thrilled to share our 📜 #paper reading schedule for #Spring 2025!
Join the list to get online login details https://t.co/rFUOHjUBNZ
Full schedule https://t.co/pddplUWRMh #Brainstorm #ProfGiri

[Ponnurangam Kumaraguru “PK” @ponguru]

🎊🥳 #PrecogTurns14 🎉🥳 Introducing new avatar of our website 😱 1st revamp in 14 yrs 😱 Curious about what keeps us awake at night? https://t.co/7bbIi58YrR Shoutout to @akshitwt for nudging. 🙏 @jswarang12 @VictorKnox99 @vnnm404 & JP garu for helping. #PrecogsRock 🚀 #ProfGiri

[Ponnurangam Kumaraguru “PK” @ponguru]

7/7 Improves decision boundaries in detecting harmful content with clearer separations between toxic and non-toxic memes.

May be of interest to @chrmanning @jurafsky @baharanm @WeiWang1973 @ZhongingAlong @sebastianseung @dmguiuc @gregd_nlp @YizhouSun

📜https://t.co/I1J7CtfvDD

[Ponnurangam Kumaraguru “PK” @ponguru]

🚀 Applications Open! 🎉 Stage 1 registrations of India AI Olympiad (INAIO), live now! https://t.co/oO4tnAe5MY Looking forward to seeing some amazing participation. Thanks 🙏🏽🙏🏽 for spreading the word! #AIOlympiad #INAIO #ACMIndia #IKDD #VPGiri