Prasoon Bajpai

Research @GoogleDeepMind | @lcs2lab |
@IITDelhi

[Prasoon Bajpai @prasNLP]

I will attend #EMNLP2024 at Miami next week! If you are interested LLM explainability, formal reasoning and/or multilingual NLP, please DM me and connect😃. I'm ready for a☕ talk every day! Also, please find me on Nov 13th 10:30-12:00 at poster session 6!

[Prasoon Bajpai @prasNLP]

Great work pushing the frontier of multimodal reasoning evals!

[Darshan Singh @ CVPR @thought2vec]

Frontier models have become excellent at understanding videos. But what happens when we test them outside the comfort zone of Western, English-centric data? In our #CVPR2026 (Highlight) work, we pushed these models to their limits to see if they can function effectively in diverse global contexts. The results? They are struggling. Work done with @NagraniArsha @skawshik11 @Harman26Singh @dinesh_tewari1 @0xtob @CordeliaSchmid Anelia Angelova @shachi_dave (1/7)

[Zi Wang, Ph.D. @ziwphd]

Check out Proactive Co-Creator on
@GoogleAIStudio , a human-AI belief alignment demo I vibe coded: https://t.co/BEcchP0K89

🧠 See & edit the AI's uncertainty via belief graph. It asks clarifying questions before creating!

📷 Try Image ➔ Story ➔ Video. You can even remix it!

[Prasoon Bajpai @prasNLP]

Interesting work from @IshaanWatts18 !

[Ishaan Watts @IshaanWatts18]

Optimize pretraining not just for loss, but for robustness to future updates.
The "best" base model does not always make the best final model.

📄 More in the paper: scaling results, Hessian analysis, and practical recipes https://t.co/yiYMIwOblM

Huge thanks to my collaborators: @CatherineL11638 @goyalsachin007 @jacspringer @AdtRaghunathan

9/9

[LCS2 Lab @lcs2lab]

🚀 #EACL2026 Sneak Peak Alert 🚀
We're excited to share a paper that we are presenting at #EACL2026 in #Morocco!
📜 Can LLMs Reason over Extended Multilingual Contexts? Towards Long-Context Evaluation Beyond Retrieval over Haystacks
👥 @AmeyHengle @prasNLP Soham Dan @Tanmoy_Chak

[Prasoon Bajpai @prasNLP]

Thrilled to see our paper accepted at AISTATS 2026!
Grateful to my co-authors, this was a fun deep dive into interpretability, control, and causal prompt edits. 🚀

[Neha Kalibhat @NehaKalibhat]

Thrilled to share that our paper on "Interpreting and Controlling Model Behavior via Constitutions for Atomic Concept Edits" has been accepted at AISTATS 2026! 🚀🚀
Read more about how input mutations can be mapped to interpretable behavioral insights.
https://t.co/iRPRJoyAso
🧵

[Prasoon Bajpai @prasNLP]

Long context multi-hop reasoning still remains a hard problem in the multilingual landscape.
Our work on relevant evals got accepted into EACL!

[LCS2 Lab @lcs2lab]

📢 #LCS2 is coming to Morocco ✈️
Happy to announce that two papers from our lab have been accepted to #EACL2026. Congratulations to all the authors, great start to the year! 🙌
#IITDelhi #EACL2026 #ACLCommunity #NLProc #AIResearch @Tanmoy_Chak

[Prasoon Bajpai @prasNLP]

Go apply!

[Prateek Jain @jainprateek_]

Thrilled to note that we are keeping the tradition of the awesome AI residency program alive in a new avatar: pre-doc researcher program at GDM-Blr -- with some amazing work done by our recent predocs including @gautham_ga_ @pranamyapk @puranjay1412 @sahilgo6801 @swaroopnath6

If you want to join this program, please apply here: https://t.co/xaL0cnv3ub

[Prasoon Bajpai @prasNLP]

@VarshitaKolipa1 true, the OTP will come with the pension

[Prasoon Bajpai @prasNLP]

@SuryaDoesIt @GoogleDeepMind Hi Surya! Usually the applications roll out near year end. Following initial screenings, there is a series of interviews and evaluations.

[Prasoon Bajpai @prasNLP]

New home at @GoogleDeepMind India as Pre-Doctoral Researcher!

[Prasoon Bajpai @prasNLP]

@partha_p_t @GoogleDeepMind Looking forward!

[Prasoon Bajpai @prasNLP]

@joykiratsingh @GoogleDeepMind Thanks! All the best to you on your journey!

[Prasoon Bajpai @prasNLP]

@ElisaKreiss @GoogleDeepMind Thank you @ElisaKreiss !

[Prasoon Bajpai @prasNLP]

@joykiratsingh @UNC @mohitban47 @uncnlp Congratulations!!

[Prasoon Bajpai @prasNLP]

10-minute thought-to-blog on 'Society of LLMs'
https://t.co/bTEZWIeXZ7

[Prasoon Bajpai @prasNLP]

Who else is smelling MCTS in the deep research blog?

https://t.co/FB1x1rPZi7

[Prasoon Bajpai @prasNLP]

NAACL 2025 🚀
Presenting “Multilingual Needle in a Haystack: Investigating Long-Context Behavior of Multilingual Large Language Models”

Paper Link : https://t.co/2OzMVvVWbT

[Tanmoy Chakraborty @Tanmoy_Chak]

Kicking off the year with a bang -- 4 papers accepted in prestigious venues this month!

#ICLR2025 -- 𝐋𝐋𝐌 𝐜𝐨𝐦𝐩𝐫𝐞𝐬𝐬𝐢𝐨𝐧:
We introduce 𝐏𝐫𝐮𝐧𝐞𝐍𝐞𝐭, a novel, dataset-free policy learning approach to model pruning, achieving high compression efficiency and performance retention, demonstrated by compressing LLaMA-2-7B with over 80% zero-shot accuracy retention at a 30% compression ratio. @iclr_conf
URL: https://t.co/dMlPbrT9Ju

#𝐍𝐀𝐀𝐂𝐋2025 -- 𝐈𝐧𝐯𝐞𝐬𝐭𝐢𝐠𝐚𝐭𝐢𝐧𝐠 𝐦𝐮𝐥𝐭𝐢𝐥𝐢𝐧𝐠𝐮𝐚𝐥 𝐥𝐨𝐧𝐠-𝐜𝐨𝐧𝐭𝐞𝐱𝐭 𝐛𝐞𝐡𝐚𝐯𝐢𝐨𝐫 𝐢𝐧 𝐋𝐋𝐌𝐬:
We introduce 𝐌𝐋𝐍𝐞𝐞𝐝𝐥𝐞, the first systematic evaluation of multilingual long-context retrieval in LLMs, revealing significant performance variations across languages and context positions, with insights to guide future evaluations. @naaclmeeting
Preprint: https://t.co/7VSCn57vxn

𝐍𝐀𝐀𝐂𝐋'25 -- 𝐂𝐨𝐮𝐧𝐭𝐞𝐫𝐬𝐩𝐞𝐞𝐜𝐡 𝐞𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐛𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤 𝐚𝐧𝐝 𝐦𝐞𝐭𝐫𝐢𝐜𝐬:
We introduce 𝐂𝐒𝐄𝐯𝐚𝐥, a dataset for evaluating counterspeech across four dimensions and a prompt-based framework using auto-calibrated CoT, offering better alignment with human judgment than traditional metrics. @naaclmeeting

𝐍𝐚𝐭𝐮𝐫𝐞 𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐈𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞:
In collaboration with AIIMS (All India Institute of Medical Sciences, New Delhi), NIMHANS, Bangalore and other NGOs, we wrote how GenAI can potentially empower multisectoral suicide prevention efforts, particularly in resource-constrained settings like India. @NatMachIntell

[Tanmoy Chakraborty @Tanmoy_Chak]

🌟 𝐀 𝐍𝐞𝐰 T𝐞𝐱𝐭𝐛𝐨𝐨𝐤 -- 𝐈𝐧𝐭𝐫𝐨𝐝𝐮𝐜𝐭𝐢𝐨𝐧 𝐭𝐨 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 🌟

I am excited to share the release of my new textbook, 𝘐𝘯𝘵𝘳𝘰𝘥𝘶𝘤𝘵𝘪𝘰𝘯 𝘵𝘰 𝘓𝘢𝘳𝘨𝘦 𝘓𝘢𝘯𝘨𝘶𝘢𝘨𝘦 𝘔𝘰𝘥𝘦𝘭𝘴 (#LLMs) -- Perhaps the first textbook on LLMs.

Target Audience:
👉 Students/beginners, Looking for a structured starting point to learn LLMs
👉 Teachers, planning to offer a course on LLMs
👉 Industry professional, seeking to deepen their understanding of LLMs

Explore the Book:
🔗 Book Website: https://t.co/tLQDfkhKOp
📑 Table of Contents: https://t.co/RyxBENqYEw
🛒 Available on Amazon: https://t.co/PX2i3OLmiG

Enhance Your Learning Experience:
👉 Slides & Lecture Videos: Chapter-wise resources -- https://t.co/9bd9lhhfAj
👉 Exercises & Solutions: Practice with detailed chapter exercises (solutions available on request).
👉 Upcoming @nptel_official Course: Starting January 2025! Preview here: https://t.co/i71bgfVne8

Book Endorsement:
📖 Foreword by Prof. Tim Baldwin @eltimster
👏 Endorsements from Prof. Iryna Gurevych @IGurevych and Prof. Pushpak Bhattacharyya

#LLMs #Textbook @iitdelhi @WileyIndiaPL @lcs2lab

[Prasoon Bajpai @prasNLP]

We also find that LLMs struggle to give proper attention to parts of queries, which are grounded in highly popular entities.
Check out the full paper for more key insights, real-world implications and detailed methodology : https://t.co/FxhQmXqHeQ

[Prasoon Bajpai @prasNLP]

“Beware the fury of the highly popular knowledge”

Does highly popular information cause any internal struggle in LLMs? (1/n)

[Prasoon Bajpai @prasNLP]

We also assess this impact critical limitation under the lens of sensitivity towards lexical variations of the queries. We unveil a key weakness in modern LLMs, in being internally sensitive to lexical perturbations, while retrieving highly popular facts from their memory.