Rakesh R Menon

CS PhD Candidate @uncnlp
@umasscs, @iitmadras alum.

[John Murzaku @jmurzaku]

built a TUI to view + debug agent trajectories with Grok Build glued in.

It caught a silent 22kb write failure the agent never verified

@grok

[John Murzaku @jmurzaku]

Infra headaches from changing docs? We built InfraReconciler at @ycombinator Hack the Stackathon.

1. scrape & snapshot months of infra docs, running daily for new changes (thanks @firecrawl and @supabase)
2. plug into your infra code on @github
3. pinpoint fixes, warnings, and new best practices with InfraAgent (thanks @openrouter).

[Sagnik Mukherjee @saagnikkk]

🚨 Paper Alert: “RL Finetunes Small Subnetworks in Large Language Models”

From DeepSeek V3 Base to DeepSeek R1 Zero, a whopping 86% of parameters were NOT updated during RL training 😮😮
And this isn’t a one-off. The pattern holds across RL algorithms and models.
🧵A Deep Dive

[John Murzaku @jmurzaku]

@PrimeIntellect I built an env to test whether audio LLMs build prosody‑only reasoning chains on MELD (multimodal emotion dataset). Models listen to an utterance (raw audio file) and then must output one sentence like:
<impression> higher pitch, noticeable variation, moderate volume, quick pace </impression>.

This was so fun I feel lots of momentum to continue this and the ideas are flowing lol.

@willcb I made some minor changes to verifiers to allow audio input (see https://t.co/iaG1J9TLlg).

HuggingFace dataset available here: https://t.co/xKj6ZxDbS0

Inspired by Jeff Wu's paper here: https://t.co/t3HHLt0T9S

Env here: https://t.co/RcN1rjWYNy

[Wadhwani School of Data Science & AI (WSAI), IITM @WSAI_IITM]

A moment of global pride for India in the field of AI.

Prof. Mitesh M. Khapra, Co-founder at the Nilekani Centre at AI4Bharat, WSAI, IIT Madras, has been featured among '2025 TIME100 AI List of the World’s Most Influential People in Artificial Intelligence'

AI4Bharat is one-of-a-kind project which collected thousands of hours of voice data across 400 districts.

His pioneering work is bridging the AI gap for Indian languages.

🔗: https://t.co/7RHGDj2ivW

@TIME @ravi_iitm @rbc_dsai_iitm @IBSE_IITM @ai4bharat @wcte_iitm @iitmadras @EduMinOfIndia @OfficialINDIAai @SarvamAI

#TIME100AI #ArtificialIntelligence #TIME100AI2025 #AI4Bharat #FutureOfAI #IndiansInSTEM #DigitalIndia #SovereignAI #IITMadras #MiteshKhapra

[Rheeya Uppaal @RUppaal]

I will be at ICLR next week to share our work on Model Editing for Alignment! DM if you'd like to chat about safety, interpretability, life in general or tourist spots in Singapore!
#NLP #AISafety #Interpretability #ICLR2025 @iclr_conf

[Somnath Basu Roy Chowdhury @SomnathBrc]

𝐇𝐨𝐰 𝐜𝐚𝐧 𝐰𝐞 𝐩𝐞𝐫𝐟𝐞𝐜𝐭𝐥𝐲 𝐞𝐫𝐚𝐬𝐞 𝐜𝐨𝐧𝐜𝐞𝐩𝐭𝐬 𝐟𝐫𝐨𝐦 𝐋𝐋𝐌𝐬?

Our method, Perfect Erasure Functions (PEF), erases concepts from LLM representations w/o parameter estimation, achieving pareto optimal erasure-utility tradeoff w/ guarantees. #AISTATS2025 🧵

[Hadas Orgad @ ICML @OrgadHadas]

🎉 Our Actionable Interpretability workshop has been accepted to #ICML2025! 🎉
&gt;&gt; Follow @ActInterp

@tal_haklay @anja_reu @mariusmosbach @sarahwiegreffe @iftenney @megamor2

Paper submission deadline: May 9th!

[Association for Computing Machinery @TheOfficialACM]

Meet the recipients of the 2024 ACM A.M. Turing Award, Andrew G. Barto and Richard S. Sutton! They are recognized for developing the conceptual and algorithmic foundations of reinforcement learning. Please join us in congratulating the two recipients! https://t.co/GrDfgzW1fL

[UMass Amherst @UMassAmherst]

Andrew G. Barto and Richard S. Sutton have been awarded the prestigious 2024 ACM A.M. #TuringAward for developing a branch of artificial intelligence known as reinforcement learning.

@UAlberta @manningcics

#ManningCICS #ArtificialIntelligence #UMass

https://t.co/x7i5v1SDUb

[Danish Pruthi @danish037]

Remember this study about how LLM generated research ideas were rated to be more novel than expert-written ones?

We find a large fraction of such LLM generated proposals (≥ 24%) to be skillfully plagiarized, bypassing inbuilt plagiarism checks and unsuspecting experts. A 🧵

[Anvesh Rao @nvshrao]

(1/4) Excited to present our latest work "𝐄𝐱𝐩𝐥𝐨𝐫𝐢𝐧𝐠 𝐒𝐚𝐟𝐞𝐭𝐲-𝐔𝐭𝐢𝐥𝐢𝐭𝐲 𝐓𝐫𝐚𝐝𝐞-𝐎𝐟𝐟𝐬 𝐢𝐧 𝐏𝐞𝐫𝐬𝐨𝐧𝐚𝐥𝐢𝐳𝐞𝐝 𝐋𝐌𝐬" at #NAACL2025! 🎉 We investigate 𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝑖𝑧𝑎𝑡𝑖𝑜𝑛 𝐵𝑖𝑎𝑠 in LLMs.

📄 Paper: https://t.co/DeA22gen8Z
🧵👇

[Nitay Calderon @NitCal]

Our paper:
"On Behalf of the Stakeholders: Trends in NLP Model Interpretability in the Era of LLMs"
has been accepted to NAACL 2025 main 🎉

We’ve updated the final version: 🔗https://t.co/m83tNeBbPi

If you are an NLP interpretability researcher, we have key takeaways for you!👇

[Rheeya Uppaal @RUppaal]

@iclr_conf paper alert! The de facto way to align a model through tuning-based methods like DPO is powerful, yet expensive and prone to jailbreaking. Emerging work on model editing aims to address this, and yet the two approaches are largely siloed. Can we somehow connect them?🧐

[IIT Madras @iitmadras]

@iitmadras has partnered with @perplexity_ai, a revolutionary search engine founded by our alumnus, Dr Aravind Srinivas (@AravSrinivas), to offer its faculty and students free access to Perplexity Pro. This cutting-edge search engine is designed to provide a more robust and versatile search experience, empowering users to conduct more thorough investigations into topics.

Perplexity Pro stands out from its free version with several enhancements, including access to more powerful AI models, a choice of various AI models for searches, and deeper search capabilities. These features enable users to delve deeper into subjects and gather more accurate information.

Prof. B Ravindran (@ravi_iitm), Head of the Wadhwani School of Data Science and AI (@WSAI_IITM) at IIT Madras, lauded this initiative, highlighting Perplexity AI's reliability and potential to transform how future generations learn and consume information online.

IIT Madras extends its heartfelt gratitude to Dr. Aravind Srinivas for this generous gesture, which will undoubtedly enhance research and learning experiences of our faculty and students.

With Perplexity Pro, IIT Madras faculty and students will have access to real-time information, conversational interfaces, source transparency, and advanced AI technology. This partnership is poised to enhance research and learning experiences, fostering a community of innovators and thinkers.

#IITMadras #PerplexityAI #ArtificialIntelligence #SearchEngine #Research #Learning #Innovation #Collaboration #EmpoweringStudents

[Ekin Akyürek @akyurekekin]

Why do we treat train and test times so differently?

Why is one “training” and the other “in-context learning”?

Just take a few gradients during test-time — a simple way to increase test time compute — and get a SoTA in ARC public validation set 61%=avg. human score! @arcprize

[Andrew Ilyas @andrew_ilyas]

Machine unlearning ("removing" training data from a trained ML model) is a hard, important problem.

Datamodel Matching (DMM): a new unlearning paradigm with strong empirical performance!

w/ @kris_georgiev1 @RoyRinberg @smsampark @shivamg_13 @aleks_madry @SethInternet (1/4)

[Kevin Ellis @ellisk_kellis]

New ARC-AGI paper
@arcprize w/ fantastic collaborators @xu3kev @HuLillian39250 @ZennaTavares @evanthebouncy @BasisOrg
For few-shot learning: better to construct a symbolic hypothesis/program, or have a neural net do it all, ala in-context learning?
https://t.co/zcmxoQzv92

[The Nobel Prize @NobelPrize]

BREAKING NEWS
The Royal Swedish Academy of Sciences has decided to award the 2024 #NobelPrize in Physics to John J. Hopfield and Geoffrey E. Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks.”

[Rheeya Uppaal @RUppaal]

Excited to share our latest research on improving the safety of LLMs! We've developed DeTox, a tuning-free and noise robust alignment method that significantly reduces model toxicity without the need for large-scale preference data. 🚀 1/n