One of the coolest things in this paper is that we now have a regression style loss function for divergence matching (just like flow matching!) which is super simple but works super well! Check out our paper and the original thread by our amazing first author @TonyRKOuYang!
Introducing SCALLOP 💜 Our new paper accepted at NeurIPS 2026 🐨
🚀 A scalable, efficient way to train few-step likelihood-available flow maps, enabling effective Boltzmann generation!
📽️ project page: tonyauyeung.github.io/scallo…
🔗 paper: arxiv.org/pdf/2606.29110
Yutong (Kelly) He retweeted
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
🧵
Yutong (Kelly) He retweeted
Can diffusion models generate high-quality images without CFG?
Yes! We introduce ✨Correction Training for Diffusion Models✨, which significantly improves sample quality without relying on classifier-free guidance.
Project page: sunovivid.github.io/correcti…
🧵 (1/n)
Yutong (Kelly) He retweeted
Introducing Gemini 4 Argon – our new frontier model.
It’s built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense – rolling out today to a set of trusted testers through our Fairwind Program.
Yutong (Kelly) He 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.
Witnessing @ssokota’s work on this project has been one of the most inspiring experiences of my PhD. It’s a reminder to me that even in this day and age, it’s ok to choose to work on the hard things, to go all in, and to pour your heart and soul into them. Huge congrats man!
Yutong (Kelly) He retweeted
The most common question we get after MaxRL is: how do you handle continuous reward?
Now I think we have a good answer to that -- check out the post from @stablegradients to see how we did it 👇
(We'll share more evidence for why we believe this is the correct extension soon.)
Is RL optimizing the right objective? 🤔
Should we maximize mean reward? Best-of-k? Which k?
Standard RL pulls on the mean and often the distribution collapses to a spike. The tail dies 🥲
We introduce Tail-Likelihood Reinforcement Learning (TailRL).
It maximizes the mean reward while simultaneously maximizing coverage over high reward outputs.
🧵 1/n
Yutong (Kelly) He retweeted
Our tech report on DiffusionGemma is out! 🚀
My biggest highlight? Online post-training is a game changer for text diffusion.
We performed joint Sampler Distillation and Reinforcement Learning (SD⋅RL) after SFT, pushing the quality-speed Pareto frontier into a new regime 📈
Yutong (Kelly) He retweeted
We just released the DiffusionGemma tech report. Text diffusion models open up a radically different part of the latency–quality Pareto frontier 👇
We hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of yet.
None of this work happened in isolation. DiffusionGemma builds on years of amazing research from the text diffusion community, and we wanted the report to recognise that. We’re standing on the shoulders of giants❤️
Let’s keep pushing towards lower-latency, higher-quality LLMs together!
Yutong (Kelly) He retweeted
As a PhD student, I was told not to work on deep RL - too full of hacks and alchemy" But after a year or two of working in this area, I’ve come to (deeply?) appreciate all of the thoughtful research that’s gone into understanding what/why things work and how to make them better.
My lab (joint with @abhishekunique7) took what we’ve learned by reading this body of literature to answer the question: what are the actual best-practices for finetuning a diffusion/flow/generative robot policy (for now, in sim)? Under one set of constraints - ample compute but limited time on your robot - @servo97 paper gives a pretty compelling answer.
Interaction with the real world is the major bottleneck in robot learning. So what would robot RL look like if we didn’t need to limit compute per interaction? Our latest work, Off-Policy Generative Policy Optimization (OGPO, accepted to ICML26) embarks on answering this question (spoiler alert: when done correctly, it helps massively!).
🧵(1/N)
Yutong (Kelly) He retweeted
Excited to share last summer's work at Google Research!
Most hybrid models today are static: each token sees the same interleaved pattern of your favorite linear model and attention.
Oryx instead varies the model used across the sequence through shared representations.
1/
Yutong (Kelly) He retweeted
Our team just released DiffusionGemma: a new state-of-the-art open diffusion LLM 🚀
Pushing the RL research for DiffusionGemma has been fascinating. RL doesn't just boost performance, it unlocks blistering speedups (>1000 tok/s on a single H100) ⚡️
DiffusionGemma is our new experimental open model with up to 4x faster output on dedicated GPUs.
Instead of predicting word-by-word, it generates entire blocks of text simultaneously. This lets the model self-correct and format complex markdown in real time.
ALT Intelligence vs Latency chart showing DiffusionGemma 26B A4B is much faster than Gemma 4 models with high intelligence.
Yutong (Kelly) He retweeted
DiffusionGemma is our new experimental open model with up to 4x faster output on dedicated GPUs.
Instead of predicting word-by-word, it generates entire blocks of text simultaneously. This lets the model self-correct and format complex markdown in real time.
ALT Intelligence vs Latency chart showing DiffusionGemma 26B A4B is much faster than Gemma 4 models with high intelligence.
Yutong (Kelly) He retweeted
An annoyingly common question I get as an AI PhD student is “When can you get the ChatGPT AI to do something useful? It can’t even work on my phone yet. Siri is pretty dumb.”
To be fair, I think their criticism is correct. I personally wished that AI was better integrated on my phone. LLMs can solve IMO problems, so shouldn’t it be a cakewalk for it to remind me of the text I forgot to respond to last week? Obviously not, since it doesn’t exist in my pocket yet. Or maybe Apple’s new update yesterday fixed this and my research project is obsolete.
We are releasing iOSWorld (iosworld.io), a dynamic iPhone benchmark with 26 newly created apps grounded in personal context. Each of the 26 apps is centrally seeded around one persona, Jordan Avery, and the apps interact together in a realistic ecosystem that reflect real app interactions. We create 133 personalized mobile agent tasks to test in this environment, and the best model, even with privileged information, only scores 51%.
Yutong (Kelly) He retweeted
Scaling laws describe how loss changes with scale. Do neurons inside models change predictably too?
We study vision and language models up to 30B params and find systematic scaling in neuron universality, specialization, and selectivity.
Paper+code: avdravid.github.io/rosetta-n…
1/n
Yutong (Kelly) He retweeted
We're presenting "Anchored Generation: Controllable Scene Variation via 3D-Aware Object Latents" at the #CVPR 4th Workshop on Generative Models for Computer Vision today!
This is preliminary work on modeling the distribution of controllable scene latents rather than pixels directly. I think this is a promising direction for generative modeling. Here, the generator predicts structured 3D latents instead of 2D layouts, allowing the model to better resolve ambiguity in object de-occlusion and replacement.
We’d be happy to discuss the work, so feel free to stop by!
🕐 Time: 1:00–2:00 PM MDT
📍 Place: Room 205, 4th Workshop on Generative Models for Computer Vision (generative-vision.github.io/…)
I’m also happy to connect with people attending CVPR. Feel free to DM or email me if you’d like to chat!
Yutong (Kelly) He retweeted
As LLMs saturate benchmarks, evaluating their five-nines reliability is crucial, but prohibitively expensive. We cut the inference cost by 5-20x on average (up to 156×) by exploiting a key insight: LLM failures are not random.
🧵[1/n]
Yutong (Kelly) He retweeted
One of the things I’m most excited about this year is building agents that can work productively for hours, days, or weeks. Coding agents are starting to become very competent at this, but what about computer use agents?
Our new benchmark, Odysseys (co-led with @JangLawrenceK) is a set of 200 new tasks derived from real world browsing behavior that measure long horizon web navigation capabilities (potentially up to hours of web browsing work). Interestingly, we find that frontier CUAs are already surprisingly good at working productively for up to an hour on these tasks, but there’s a lot of work to be done in making them even more efficient.
Like every other AI researcher, my real dream is to open a cafe once we solve ASI. So, here’s Opus 4.6 doing some market research for me ("I want to do market research on the most popular cafes in Singapore. Analyse the menus of the top 10 cafes in Singapore (by Google reviews/ratings), and make sure we include at least 1 from the North/South/East/West/Central regions of Singapore. Keep the relevant pages of each cafe open, and summarise their pricing, menu offerings, unique selling points, making sure to reference which tab is opened for each cafe. For each cafe, also help me figure out how long it would take to get to it from Tampines MRT, and include this in your final summary.").
I was very impressed to see Opus 4.6 complete this task after working for 52 mins, satisfying all 7 rubrics that corresponded to this task. It provided a very nice markdown summary at the end that gave me all the information I asked for!
Yutong (Kelly) He retweeted
Foresight will be the defining frontier on the path to AGI.
I am excited to start Sooth Labs with my amazing co-founders: Yaser Sheikh @subail, Chuck Hoover @chuckjhoover, David LaRose, and Shih-En Wei.
Deeply grateful to Aydin Senkut @asenkut and Feyza Haskaraman @FHaskaraman at @felicis for leading the round, alongside an exceptional group of partners.
bloomberg.com/news/articles/…
Diffusion planners are great for offline RL. But they need many steps to work well! Way too slow for real-time decision making!
Presenting RACTD at #ICLR2026: reward-aware distillation that plans in ONE step
🇧🇷 Today (4/23) P4-#4618 3:15-5:45 PM
arxiv.org/abs/2506.07822 1/
RACTD improves over previous SOTA by 9.7% on D4RL Gym-MuJoCo and outperforms Diffuser on long-horizon Maze2D planning
All with a SINGLE denoising step, achieving up to 142x speedup over diffusion counterparts 🚀🚀🚀
Joint work with the amazing @x_duan9296 (the best master's student at CMU), @FahimTajwar10, @rsalakhu, @zicokolter, and Jeff Schneider. Come chat with us today at 3:15 PM!
Paper: arxiv.org/abs/2506.07822
Code: github.com/Ddduanxt/RACTD
