Past: Research @GoogleDeepMind (Gemini 3) | @iiit_hyderabad
Montréal
Joined October 2021
- Tweets68
- Following526
- Followers209
- Likes2.7K
Check out this amazing work by @Harman26Singh and team @AIatMeta! Traditionally, agents are evaluated on how good they are at a task. But as agents run longer, an equally important question is how efficiently they improve with experience (plasticity). Turns out capability and plasticity are pretty different things.
Darshan Singh retweeted
Great work, important to benchmark on the right things! Happy to see that VELOCITI, our CVPR 2025 work survives! 🎉
Try it out, short 10 second videos that stump many models.
katha-ai.github.io/projects/…
Darshan Singh retweeted
Excited to be at ICML'26 🇰🇷 from July 6 onward!
I'll be presenting papers on Reasoning, RL and Diffusion Language Models.
Would love to chat about RL for agents, text-optimization, self-improvement, and diffusion LMs.
See you in Seoul 🚀
Paper threads and details below 👇
Darshan Singh retweeted
Check out this new work by Darshan and gang! Really enjoyed working on non-western, non-english domains with such an amazing team. At CVPR this year!
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)
Super thrilled to share that our work, SRL-CLIP, won the Best Paper Award at the CV4Smalls #CVPR2026 workshop! 🏆 Huge thanks to my amazing co-authors @zeeshank95 and @MakarandTapaswi .
Also, a big thank you to the organizers (@shaydamoezzi and team) for putting together such a thoughtful workshop. With the current trend of "scaling everything," it is incredibly refreshing to focus on data inefficiency as a design opportunity rather than a limitation. Data efficiency is one of the main pillars needed to improve real-world video understanding.
ALT Best paper award certificate
Hi all! Today I will present our work, SRL-CLIP, a recipe to efficiently adapt CLIP to videos at the CV4Smalls CVPR workshop in Room 102. I have an oral talk scheduled at 2 pm followed by the poster session. Please do stop by! Work done with @zeeshank95 @MakarandTapaswi #CVPR2026
ALT The poster of our paper on efficient CLIP adaptation for videos
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)
This allows us to independently assess each intermediate step of the model's thought process. 🔍 Through our graph-based error tracing, we discovered that roughly 75% of all model failures actually stem from the visual perception of cultural elements. 🤯(6/7)
MINERVA-Cultural is a step toward helping the world build more equitable and diverse VLMs 🌍. Huge thanks to all my co-authors and the team!. Check out the paper here: arxiv.org/abs/2601.10649 and the dataset here: github.com/google-deepmind/n… #CVPR2026 #AI #ComputerVision (7/7)
Hi all! Today I will present our work, SRL-CLIP, a recipe to efficiently adapt CLIP to videos at the CV4Smalls CVPR workshop in Room 102. I have an oral talk scheduled at 2 pm followed by the poster session. Please do stop by! Work done with @zeeshank95 @MakarandTapaswi #CVPR2026
ALT The poster of our paper on efficient CLIP adaptation for videos
Grateful to be recognized as an Outstanding reviewer at #CVPR2026. Always happy to do my part and give back to the community. This CVPR is very special for me because we also have 1 Main paper (Highlight) and 1 workshop paper (Best paper candidate). See you all in Denver! 🚀
We are grateful to all of the 17,491 reviewers who helped make #CVPR2026 possible. We are especially pleased to recognize the following Outstanding Reviewers, whose high-quality reviews (as judged by their Area Chairs) placed them among the top 5% of reviewers.
Darshan Singh retweeted
The Project Aria team is excited to be part of the Third Joint Egocentric Vision (EgoVis) Workshop at #CVPR2026! 👓✨
Don’t miss the latest Project Aria updates as we dive into the future of egocentric perception. Check out the full program of invited speakers below!
📅 June 3, 2026
📍 Colorado Convention Center | Room 704/706
🔗 egovis.github.io/cvpr26/
⚡️@liuziwei7 , @mapo1 , @doughty_hazel
Darshan Singh retweeted
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
Darshan Singh retweeted
Pretrained ViTs like DINOv2 or CLIP are great, but they produce fixed, generic representations that encode the most salient visual concepts (e.g., "cat").
In human vision, prior priming with language changes how people parse an image. We believe visual encoders should do the same
🚨 Introducing Steerable Visual Representations, a new family of visual features you can steer with text towards specific visual concepts.
Darshan Singh retweeted
Can LLMs Self-Verify? Much better than you'd expect.
LLMs are increasingly used as parallel reasoners, sampling many solutions at once.
Choosing the right answer is the real bottleneck.
We show that pairwise self-verification is a powerful primitive.
Introducing V1, a framework that unifies generation and self-verification:
💡 Pairwise self-verification beats pointwise scoring, improving test-time scaling
💡 V1-Infer: Efficient tournament-style ranking that improves self-verification
💡 V1-PairRL: RL training where generation and verification co-evolve for developing better self-verifiers
🧵👇
Darshan Singh retweeted
Our paper was desk rejected @NeurIPSConf! Even before the main deadline! "Non-academic title and abstract" 🙈
Thankfully, @SIGGRAPHAsia was around the corner and a perfect fit for our work on improving robustness of multi-subject multi-attribute layout-guided T2I models!
🧵1/9