Bhuvan Sachdeva

Visiting Researcher at @MSFTResearch | Past Research at @AmazonScience
Looking for Research Engineer roles!

[Bhuvan Sachdeva @SachdevaBhuvan]

Our paper has been accepted for Oral presentation at #CVPR 🎉🎉
Kudos to the team: @karan_uppal3, @abhinav_java
See you in Denver!

On a side note, I am finishing my fellowship by June and looking for full-time research roles. DMs open.

[Bhuvan Sachdeva @SachdevaBhuvan]

New paper: Understanding Task Transfer in Vision–Language Models
How does finetuning a model on one task affect its performance on other tasks?
@karan_uppal3 and @abhinav_java are presenting this work at Unireps, NeurIPS!!
📍 Ballroom 20D
⏰ 3:45 PM – 5:00 PM
Come and say Hi!🧵

[Bhuvan Sachdeva @SachdevaBhuvan]

A very thorough and refreshing take on AI research.

[Jack Morris @jxmnop]

https://t.co/w8pDpes7zL

[Karan Uppal @karan_uppal3]

Excited to be heading to #CVPR2026 to give an oral presentation of our paper on task transfer!

If you're in Denver, do attend the presentation on Sunday. Looking forward to meeting fellow researchers!

Link: https://t.co/N65Mkvl0km

Special thanks: @SachdevaBhuvan @abhinav_java

[Bhuvan Sachdeva @SachdevaBhuvan]

Checkout our latest work!!

[Deepti Ghadiyaram @deeptigp]

[1/4] The human eye doesn't process every single pixel of a video continuously—it focuses on what changes.

So why are our video AI models wasting compute on redundant frames?

Introducing Swift Sampling: a test-time technique inspired by the human visual system. 🧠👇

[Bhuvan Sachdeva @SachdevaBhuvan]

@gabriberton I agree with the pros for vision encoder and think vision encoder are here to stay. That being said, do you think there's something inherently special with the visual modality to require separate encoding or are the advangates merely compute efficiency?

[Bhuvan Sachdeva @SachdevaBhuvan]

Checkout the exciting work by Ishaan!

[Ishaan Watts (ICML’26 🇰🇷) @IshaanWatts18]

Spending billions to train the "best" base model? You might be optimizing the wrong thing! 🎯

We show that controlling sharpness during mid-training leads to over 35% less forgetting after fine-tuning / quantization... even when the base model itself gets worse.

🧵 Takeaways for pretraining:
- Use SAM (Sharpness-Aware-Minimization) in the final steps (~10%)
- Try much higher learning rates (yes, even ~10× larger)

1/9

[Saining Xie @sainingxie]

vision🍌 is here https://t.co/Ued6GGk4Et
if you got into computer vision the way I did, starting with pixel-level labeling tasks like segmentation, edges, depth, or surface normals, you’ll probably feel the same seeing these results -- something big has quietly shifted, and it’s going to change how we approach these problems for good 🧵

[Bhuvan Sachdeva @SachdevaBhuvan]

@dhruvverma_ Crazy, congrats!!

[Nalin Wadhwa @ ICLR 2026 @nalin_wadhwa]

Excited to announce that I will be giving an oral presentation of BEAVER at the VerifAI workshop @ ICLR 2026!
Come chat if you are interested in formal guarantees for LLMs 🦫

[André Araujo @andrefaraujo]

True multimodal AI needs to understand the world spatially 🎯
🚀 Excited to release #CVPR2026 TIPSv2 from @GoogleDeepMind, a foundational image-text encoder with spatial awareness, leading to strong overall results and massive gains on patch-text alignment. 🔥
1/N

[Lorenzo Xiao @lrzneedresearch]

Made a public RL-for-LLMs reading list because I was trying to prepare for my interviews

96 papers, 5 categories, 24 subtopics, mostly around the 2025-2026 wave, with notes on what’s worth reading carefully vs what you can skim.

Hopefully useful if you’re getting into RLHF/agentic RL/ reward modeling, or just cramming for interviews.

https://t.co/DquWjcKmCb

Note: Paper selection mainly adhere to me and @sun_hanchi's taste...

Make sure you follow me so I can have the incentive to update this and the agentic system design series

#RLHF #LLM #AIAgents

[Bhuvan Sachdeva @SachdevaBhuvan]

@_Creation22 Does a given sequence have a single solution only?
For example, take a sequence that has all the numbers in order from 1 to 25, except 12.
You can use 12 from the start and decompose 21 into 1 and 2. This way, the sequence can be missing 12 or 21.

[Bhuvan Sachdeva @SachdevaBhuvan]

Pretty good read, it reminded me of "You and Your Research" by Richard Hamming (which also referred in the article itself).

[Nathan Lambert @natolambert]

Good post on how to think about honing your skills as an (academic) researcher by Carlini

https://t.co/pT5XT6R2oK

[Bhuvan Sachdeva @SachdevaBhuvan]

Full paper accepted to CVPR!!😀

[Bhuvan Sachdeva @SachdevaBhuvan]

New paper: Understanding Task Transfer in Vision–Language Models
How does finetuning a model on one task affect its performance on other tasks?
@karan_uppal3 and @abhinav_java are presenting this work at Unireps, NeurIPS!!
📍 Ballroom 20D
⏰ 3:45 PM – 5:00 PM
Come and say Hi!🧵

[Bhuvan Sachdeva @SachdevaBhuvan]

Checkout Nalin's first PhD project!!

[Nalin Wadhwa @ ICLR 2026 @nalin_wadhwa]

🚨 Can we truly prove that LLMs won't leak your data or generate unsafe code?

Almost all LLM safety evals today only give estimates.
Sampling. Benchmarks. Leaderboards.
Zero guarantees ❌

We present BEAVER, the first system to provably bound LLM constraint satisfaction.

🧵👇

[Bhuvan Sachdeva @SachdevaBhuvan]

AI generated references about to go insane!

[OpenAI @OpenAI]

Introducing Prism, a free workspace for scientists to write and collaborate on research, powered by GPT-5.2.

Available today to anyone with a ChatGPT personal account: https://t.co/9mTLAbxPdH

[Bhuvan Sachdeva @SachdevaBhuvan]

(5/5) For more details, check out the paper: https://t.co/XPEj1rwNqw. Happy to answer questions! Shoutout to my amazing co-authors and mentor Vineeth N B.

[Bhuvan Sachdeva @SachdevaBhuvan]

New paper: Understanding Task Transfer in Vision–Language Models
How does finetuning a model on one task affect its performance on other tasks?
@karan_uppal3 and @abhinav_java are presenting this work at Unireps, NeurIPS!!
📍 Ballroom 20D
⏰ 3:45 PM – 5:00 PM
Come and say Hi!🧵

[Bhuvan Sachdeva @SachdevaBhuvan]

(4/5) Why use PGF? PGF-guided selection enables alternative datasets that rival or surpass direct finetuning when supervision data is scarce.