phd @berkeley_ai, prev iitd

Berkeley, CA
Joined June 2019
Himanshu Gaurav Singh retweeted
Introducing ✨Token-Gated Language Models✨ for up to 11x faster prefill by learning to skim context, conditionally forwarding only tokens that matter. Akin to the shift from dense models to MoEs, the next frontier of efficient LLMs is conditional computation over a new scaling axis: context tokens🧵👇
29
64
13
361
43,513
Himanshu Gaurav Singh retweeted
Two Sunday reads
8
2
60
4,818
REINFORCE is susceptible to getting stuck in local optima of the reward landscape. GracefulRL investigates this and provides a principled fix! Check out @damanarora’s thread for more details
1/ RL for improving task success has made LLMs remarkably capable. But what happens when the model fails? Turns out, training for graceful failure can be surprisingly tricky! Very excited to share our new paper: Training Language Models to Fail Gracefully 🧵
4
21
1,163
Himanshu Gaurav Singh retweeted
The problem being identified here has nothing to do with mathematics, clinical trials, or human understanding. The problem is that we have mechanisms for assigning (and enforcing) legal, social, and moral responsibility in humans, but not in AIs. This problem is solvable.
This is a slide from Terence Tao's 'Math 2.0' public lecture yesterday evening at Caltech. I am highly confident that, if the cure actually worked, almost any Stage 4 cancer patient would tell you that human comprehension of its mechanism is completely irrelevant to them.
22
9
3
182
12,568
Himanshu Gaurav Singh retweeted
(1/5) Everyone says dexterous manipulation needs real-world data. These policies were trained with none. Our robots mesh gears, thread nuts, and insert pegs zero-shot from RGB, and the same recipe works for locomotion across terrains much harder than before. The missing piece wasn't a better RL algorithm, but a small tweak to what the robot practices. Today we're releasing Success-Guided Sampling (SGS), a simple change to the task sampler that unlocks new levels of dexterity and agility with large-scale, end-to-end sim2real RL. No demonstrations, no per-task reward tuning. Just PPO at scale. Check out the policies yourself at sgs-rl.github.io! All videos at 1x speed! 🧵 🔈
43
185
70
1,075
206,493
Himanshu Gaurav Singh retweeted
Please resist the urge to mock the mathematicians struggling through this moment. Whatever your field, your expertise, your passion, you too shall one day experience what they are going through. And when you remember your words, you will feel remorse.
696
422
222
5,342
372,096
Systematic measurement of self improvement is crucial before actually achieving it!
🤖 How well do agents learn from experience? We study agents that amortize experience into reusable artifacts and introduce Agent Plasticity to measure how efficiently they improve on held-out environments. Top performer ≠ most efficient learner.
2
1
22
1,244
Himanshu Gaurav Singh retweeted
Can “chicken” make a base model reason? 🐔 Yes! Why? With the right first tokens, base models can match RL in reasoning. We trace this to learned training data associations and use a simple data edit to make “chicken” elicit reasoning too. blog: sophielwang.com/cues
48
80
29
615
194,925
Statement : The removal of BitChat from the India App Store is unconstitutional. Section 69A permits the blocking of unlawful information, not of a messenger for working during internet shutdowns. New Delhi, 3 October 2026 An Apple App Review notice published today by Jack Dorsey on X records that BitChat has been removed from the India App Store "per demand from the Ministry of Electronics and Information Technology", issued under Section 69A of the Information Technology Act, 2000, on the ground that the application "includes content that is illegal in India". The notice adds that BitChat's TestFlight builds are no longer available for testing in India. It does not identify the content said to be illegal or the ground under Section 69A invoked. The blocking order has not been published. The Internet Freedom Foundation condemns the order and calls for its immediate publication and withdrawal. BitChat relays end-to-end encrypted messages between nearby phones over Bluetooth mesh. It has no servers and hosts no content of its own. What travels through it is the private speech of its users, so the claim that the application "includes" illegal content is incoherent. The State's actual objection stands recorded in the notices the Indian Cyber Crime Coordination Centre issued on 23 July 2026 to GitHub, Google and Apple against BitChat, Briar and Bridgefy which stated that such applications work even during internet shutdowns and hamper lawful interception, attribution and traceability. Those notices, issued under Section 79(3)(b) were reportedly withdrawn after public disclosure. The same end has now been pursued through Section 69A. The objection is to capability of an application, not to content it transmits. A citizen's ability to speak to another citizen while the State has switched the internet off is not a ground of restriction under Section 69A or Article 19(2) of the Constitution. Section 69A survived constitutional scrutiny in Shreya Singhal v. Union of India (2015) because of its safeguards which included the need for a reasoned, recorded order, committee review, and a hearing to the originator under Rule 8 of the Blocking Rules, 2009, which the Delhi High Court enforced in Tanul Thakur v. Union of India (2022). Further, Anuradha Bhasin v. Union of India (2020) requires that orders restricting fundamental rights be published and adopt the least restrictive measure. However, it appears in this instance, each safeguard has been evaded. A secret order blocking an entire application, rather than any identified unlawful information, is the most restrictive and disproprotionate measure available. Further, the confidentiality claimed under Rule 16 cannot defeat the right to challenge an order, and the constitutionality of this blocking regime is itself pending consideration before the Supreme Court. While removal from the App Store does not delete BitChat from phones, it stops new downloads and security updates, leaving existing users on unpatched versions, and it ends security testing in India through TestFlight. During the Jantar Mantar protests, when mobile internet was suspended and jammers were deployed, Bluetooth mesh messaging was a channel of last resort through which young people coordinated their safety. To our knowledge, Apple has removed BitChat on a government's demand in only one other country. This is in China, in April this year. Our demands 1. MeitY must publish the blocking order, the ground invoked and the record of the committee's proceedings, and serve the developers with a copy and an opportunity of hearing. 2. The direction must be withdrawn. Any specific unlawful use must be met with targeted proceedings under due process, not the removal of a communications tool for everyone. 3. Apple must disclose the demand in its transparency reporting and notify its users in India, and must state whether parallel directions have been received in respect of other app stores. 4. The secrecy of the blocking regime must end. The Blocking Rules, 2009 should be amended to require publication of orders and notice to originators, without awaiting the outcome of the pending constitutional challenge. The Internet Freedom Foundation will file applications under the Right to Information Act seeking a copy of the order, and stands ready to assist the developers and affected users in India.
185
908
56
2,952
165,189
Himanshu Gaurav Singh retweeted
this sets a terrible precedent. no government has the right to limit free speech. we don't have to accept this. we have the technology to give back power to the people. let's use it.
government of India officially removes bitchat from the App Store
57
147
11
1,000
24,144
Himanshu Gaurav Singh retweeted
1930 Cartoon by Kesava Shankar Pillai #Gandhi
5
274
3
1,099
12,468
Himanshu Gaurav Singh retweeted
The performance of a real-world robotics problem is likely never dominated by numeric. What if the arm starts moving? You can crank up the gain, or argue for inverse dynamics. But what if you have friction, inertia mismatch, state-full friction, and sensor noise? Disturbance observer and adaptive controller running online are all under strong assumptions and will not scale. If we can effectively simulate it, why don’t we solve it offline, regardless of whether it’s a humanoid or a tabletop 6DOF arm?
2
9
2
59
10,507
Himanshu Gaurav Singh retweeted
If you’re at IROS2026 @Cinnabar233 is giving a talk about ABC at the open robotics software tomorrow 10/1 at 9:35am, room 302!! sites.google.com/rice.edu/ir…
ABC full release! Full stack of goodies in here. Everything you need to get started. Here's a challenge: Can you get the robot to fold a paper airplane? Excited to see what cool things you come up with!
3
24
4,536
Himanshu Gaurav Singh retweeted
The key to continual learning is through **efficient** and **generalizable** deployment-time updates. How do we teach the model this important meta-capability without complicated training? Check out @ZAlvin39105's post to see how we achieve this via **Harness Learning** 👇
1/8 How do we train a model to adapt to tasks it has never seen, without updating its weights at test time? With Harness Learning, we train a model to revise harnesses using execution feedback. At test time, the model applies this learned skill to improve harnesses on unseen tasks, while model weights stay fixed. Improving agent harnesses takes experimentation. We want a model to learn from those revisions and make better edits on unseen tasks. Models trained with harness learning nearly double their performance on challenging unseen reasoning and multihop question-answering tasks.
8
35
3,366
Himanshu Gaurav Singh retweeted
The CEC is first among equals, not a supremo. The Indian Express reports published today on the functioning of the EC make it clear that the CEC took unauthorised and illegal steps. Gyanesh Kumar should take one more step: step down. Now.
213
437
72
2,886
201,560
Himanshu Gaurav Singh retweeted
🚀 Looking for a reliable Franka controller with seamless sim-to-real transfer? Meet FrankaTwin: github.com/tsrobcvai/frankat… 🎮 Joint/Cartesian position & impedance + gripper 🎯 System ID: 3.6 mm EE / 28 mrad joint RMSE (sim vs. real)
2
7
2
57
13,165
Hiking on the weekend ❌ Pacing the frontier ✅
1
6
53
3,311
Himanshu Gaurav Singh retweeted
Congratulations on the full release!! Now you can just deploy a capable bimanual setup anywhere for under 6k with the discount code “abc.bot”. Limited time only!
Happy to announce the full ABC release! We’re also excited that ABC was accepted to CoRL 2026! Check out our website for code, 400+ hours of sim data on 24 tasks, and 5,850 labeled policy-evaluation episodes. @arthurallshire @Cinnabar233 @ritvik_singh9 @redstone_hong @davidrmcall
1
7
149
23,286
Himanshu Gaurav Singh retweeted
Full ABC release 🔤 ! Train, test, and evaluate your models. The full suite of simulation tasks and data has also been released! Excited to see what people build 🔨🤖
Happy to announce the full ABC release! We’re also excited that ABC was accepted to CoRL 2026! Check out our website for code, 400+ hours of sim data on 24 tasks, and 5,850 labeled policy-evaluation episodes. @arthurallshire @Cinnabar233 @ritvik_singh9 @redstone_hong @davidrmcall
4
6
2
48
8,136
ABC code and policies are out!! If you want to build your robot workstation for trying out our policies, abc.bot/hardware.html provides a guide!
Happy to announce the full ABC release! We’re also excited that ABC was accepted to CoRL 2026! Check out our website for code, 400+ hours of sim data on 24 tasks, and 5,850 labeled policy-evaluation episodes. @arthurallshire @Cinnabar233 @ritvik_singh9 @redstone_hong @davidrmcall
1
41
3,586