PhD Student at @MIT_CSAIL working on data-centric AI
Joined September 2020
- Tweets47
- Following578
- Followers367
- Likes1.6K
Neha Hulkund retweeted
Marin 535B-A23B is ~7% done training, and so far things look on track. Next Tuesday (Sept 1 @ 10 PT), we will have a Zoom panel/discussion where the Marin team will talk about the design decisions that went into this run, the tradeoffs made, and our learnings. If you're interested, join the Discord for details (discord.gg/J9CTk7pqcM).
Neha Hulkund retweeted
🚢 Marin 535B-A23B started training this week! As usual, the whole process is open.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow.
Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.
Neha Hulkund retweeted
I'll be at #ICML2026 to present our *spotlight* work! TLDR: LLMs can learn to self-generate curricula for problems they can't yet solve, using self-play with meta-RL.
Please reach out to chat about self-improving agents, synthetic data & environments, curriculum learning, or anything else!
We've updated the paper with some fun additions ⬇️
Can a model learn to break its own reasoning plateau?
In our new paper, we show that LLMs can be taught with meta-RL to generate their own "stepping stones" that kickstart learning on hard math problems (0/128 success rate) where direct RL fails.
Paper 📝: arxiv.org/abs/2601.18778
Blog post 🌐: ssundaram21.github.io/soar/
(1/n)
Neha Hulkund retweeted
Excited to share ID-Sim, our identity-focused similarity metric, presenting at #CVPR2026 this week in Denver! 🎉
Humans are remarkably good at distinguishing highly similar objects across different contexts.
We asked: can we train a metric that does the same?
Neha Hulkund retweeted
As a geometric ML researcher, I noticed pseudoscalars don’t get enough attention! Read on to see what pseudoscalars can do for you.
Neha Hulkund retweeted
LLMs can learn to self-generate curricula for hard problems that they can't yet solve! Using meta-RL, with rewards grounded in learning progress, models produce their own stepping stones that kickstart learning on hard problems where direct RL plateaus.
Poster at the ICLR RSI workshop today!
Can a model learn to break its own reasoning plateau?
In our new paper, we show that LLMs can be taught with meta-RL to generate their own "stepping stones" that kickstart learning on hard math problems (0/128 success rate) where direct RL fails.
Paper 📝: arxiv.org/abs/2601.18778
Blog post 🌐: ssundaram21.github.io/soar/
(1/n)
Neha Hulkund retweeted
NVFP4 allows models to be quantized to 4 bits without too much performance degradation, but can we push 4-bit performance even further?
Today, we're releasing a new class of low-precision block-scaled data types that natively adapt to your input data: for 4-bit quantization, IF4 (Int/Float 4) allows each scaled group of 16 values to be saved as FP4 or INT4 depending on which option offers less error. Selections are recorded using the scale factor’s sign bit, which is unused in NVFP4, allowing IF4 to offer better performance with no memory overhead!
Our data types provide better downstream accuracy in LLMs, they can be implemented efficiently in next-generation hardware accelerators, and they reveal some interesting insights about low-bit quantization! 🧵
Neha Hulkund retweeted
Targeted instruction tuning for LLMs involves selecting a subset of instructions from a candidate pool using a small query set from target tasks. Despite growing interest, we still lack guidance on what to select. Our new preprint brings clarity to this space (thread 👇).
Neha Hulkund retweeted
I am on the job market this year! My research advances methods for reliable machine learning from real-world data, with a focus on healthcare. Happy to chat if this is of interest to you or your department/team.
Neha Hulkund retweeted
So excited to announce the DCVLR (Data Curation for Vision-Language Reasoning) competition at NeurIPS 2025, led by @Oumi_PBC and sponsored by @LambdaAPI!
🌟open-data 🌟
🤖 open-models 🤖
💻 open-source 💻
💪anyone can compete for free 💪
dcvlr-neurips.github.io/
🧵 1 / n
Neha Hulkund retweeted
The submission portal is now OPEN to take part in this interesting @NeurIPSConf 2025 data curation competition!! This is the first open-data, open-models, open-source competition for data curation in vision-language reasoning -- learn more 👇
dcvlr-neurips.github.io/
So excited to announce the DCVLR (Data Curation for Vision-Language Reasoning) competition at NeurIPS 2025, led by @Oumi_PBC and sponsored by @LambdaAPI!
🌟open-data 🌟
🤖 open-models 🤖
💻 open-source 💻
💪anyone can compete for free 💪
dcvlr-neurips.github.io/
🧵 1 / n
Neha Hulkund retweeted
If you are attending #ICML2025, check out our DataWorld workshop on Sat July 19. We have updated the website with more info on speakers & accepted papers! dataworldicml2025.github.io/
Also happy to chat offline about all things ✨ data ✨
Neha Hulkund retweeted
📢 Announcing our data-centric workshop at ICML 2025 on unifying data curation frameworks across domains!
📅 Deadline: May 24, AoE
🔗 Website: dataworldicml2025.github.io/
We have an amazing lineup of speakers + panelists from various institutions and application areas.
Neha Hulkund retweeted
What happens when models see the world as humans do?
In our #NeurIPS2024 paper we show that aligning to human perceptual preferences can *improve* general-purpose representations!
📝: arxiv.org/abs/2410.10817
🌐: percep-align.github.io
💻: github.com/ssundaram21/dream…
(1/n)
Neha Hulkund retweeted
The BeeryLab is busy at @eccvconf today!
@juliachae_ and @EdwardVendrow co-organized @CV4E_ECCV workshop, happening 8:30-1, where I'll moderate the panel and
Kai van Brunt and @__justinkay will present fish counting in sonar and Jae Joong Lee will present 3D trees
... – at Milan, Lombardy
Neha Hulkund retweeted
Synthetic data has huge potential to drive new improvements in training and evaluation for computer vision. Interested in learning more about advancements and challenges? Join us at the SynData4CV Workshop at #CVPR2024 tomorrow (June 18)!
syndata4cv.github.io
Calling all #CVPR2024 attendees!
Join us at the SynData4CV Workshop at @CVPR (Jun 18 full day at Summit 423-425) to learn more about recent advancements in synthetic data!
Explore more: syndata4cv.github.io/
Neha Hulkund retweeted
@CVPR AI for Conservation happy hour on Monday! Open to anyone working on or interested in the intersection of CV/ML and ecology, conservation, sustainability, climate, etc
Hosted w/ @sarameghanbeery @timmhaucke
Neha Hulkund retweeted
Delighted to share one of my favorite pieces of work, now published: PURPLE, a method to estimate disparities in the prevalence of underreported outcomes, in women’s health and healthcare more broadly.
nature.com/articles/s44294-0…
Neha Hulkund retweeted
Had a fun time attending NYC Computer Vision Day + sharing my work on Amodal Completion via Progressive Mixed Context Diffusion #CVPR2024
Project page: k8xu.github.io/amodal/
Event: cs.nyu.edu/~fouhey/NYCVision…
Thank you to the organizers and sponsors! 🗽
