Matthew Yang retweeted
We found a way to scale humanoid control that reduces GPU-hours by more than 1000ร.
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Project: supermimic.github.io
Demo: supermimic.github.io/demo/
Matthew Yang retweeted
How about having some fun with this "๐๐จ๐จ๐ฉ๐๐ ๐๐จ๐๐๐ฅ ๐๐ญ๐ฅ๐๐ฌ"? ๐
Fold depth into temporal recurrence, watch connections flow like a magic circle, and see how each paper handles its KV cache.
Safe travels! Don't get lost in the Sea of Recurrence ๐
Link below ๐
Matthew Yang retweeted
We're open-sourcing RL360, the RL infrastructure behind K2 Horizon's six domain experts: 8,192 concurrent sandboxes, 150 steps in 7 days, ~1.3M rollouts.
Long-horizon agentic RL runs thousands of agents in real sandboxes on tasks that take minutes. Some of those sandboxes crash mid-task, and a crash arrives as the same low score as a wrong answer.
Without infrastructure that can keep up, GPUs are underutilized and the model gets penalized for system crashes rather than its own mistakes.
RL360 is built for this. We share how it works, what broke at scale, and what we learned in the thread.
Code: github.com/ifm-ai/RL360
Run: wandb.ai/mbzuai-llm/rl360-puโฆ
Matthew Yang retweeted
๐ Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. Weโre sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
๐ github.com/deepseek-ai/DeepIโฆ
1/4
Matthew Yang retweeted
1/11 Can a single-token importance weight preserve the target of full-sequence correction? With Future-KL, we prove that both updates share the same unique stationary policy. We also contribute the promising empirical results we observed.
arxiv.org/abs/2610.18738
Matthew Yang retweeted
Surprising new results:
Giving tools to LLMs makes them less able to say NO!
We tested 11 multimodal LLMs (Claude, GPT, Gemini, ...). With tools, every one fails to refuse harmful requests more often, up to 68.7% more!
Details and possible reasons ๐งต
To appear at NeurIPSโ26.
Matthew Yang retweeted
How much do LLMsโ watermarks distort your output? ๐
We present SimpleMark: a fast multi-bit text watermark that bounds the distortion of every token to a chosen f-divergence budget. Our watermarkโs text distribution stays statistically close to what the model would have written without a watermark while remaining competitive in all benchmarked performance categories.
Matthew Yang retweeted
Weโre releasing a broad range of new mathematical results produced by an internal frontier model.
Weโve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
Matthew Yang retweeted
Is Jev secretly learning a Value function?
A model can always choose โAโ without being certain that โAโ is correct. That distinction matters when an agent must decide whether to act fast or switch to deeper reasoning.
We spent some time exploring:
1. How Jev fits into System 1/2 decision-making and why calibration matters for knowing when to switch
2. Why maximising expected reward in RL does not guarantee calibrated probabilities.
3. How predicting eventual task success in agentic tasks connects to a familiar RL quantity: the Q-function.
Could value prediction help build a calibrated System 1โand could that be part of the story behind Jev?
Link: souradip-chakraborty.github.โฆ
#Jev #RLCD #System1
Matthew Yang retweeted
SFT is not dead! ๐ฅณ
We found a way to make SFT rival current prevailing posttraining methods, often generalizing better and forgetting less than RL and OPSD. ๐คฏ
Following our prior work on reasoning with sampling, we now introduce sampling to the posttraining stack.
1/n
Matthew Yang retweeted
๐ ๐๐ก๐๐ญ ๐ข๐ ๐ฅ๐จ๐จ๐ฉ๐๐ ๐๐๐ฌ ๐๐จ๐ฎ๐ฅ๐ ๐ฎ๐ฌ๐ ๐ร ๐๐๐ฐ๐๐ซ ๐ฉ๐๐ซ๐๐ฆ๐๐ญ๐๐ซ๐ฌ ๐๐ง๐ ๐ ๐ร ๐ฌ๐ฆ๐๐ฅ๐ฅ๐๐ซ ๐๐ ๐๐๐๐ก๐ ๐๐ง๐ ๐ฌ๐ญ๐ข๐ฅ๐ฅ ๐๐ ๐๐จ๐ฆ๐ฉ๐๐ซ๐๐๐ฅ๐ ๐ญ๐จ ๐ ๐ฌ๐ญ๐๐ง๐๐๐ซ๐ ๐๐ซ๐๐ง๐ฌ๐๐จ๐ซ๐ฆ๐๐ซ?
Introducing ๐๐๐ญ๐ก๐ข๐ง๐ค๐ข๐ง๐ ๐๐ญ ๐
๐ข๐ฑ๐๐ ๐๐จ๐ข๐ง๐ญ๐ฌ. We argue that looped LMs don't have to pay for every loop. They can scale by adding FLOPs at constant memory.
The key is the fixed point of the loop. We introduce:
โ ๐
๐ข๐ฑ๐๐-๐๐จ๐ข๐ง๐ญ ๐๐ก๐จ๐ซ๐ญ๐๐ฎ๐ญ๐ฌ: what fixed points buy a looped model in training (pre-training, post-training) and inference (prefill, decoding)
โ A ๐ฅ๐๐๐ซ๐ง๐๐ ๐๐๐ฉ๐ญ๐ก ๐ฉ๐ซ๐ข๐จ๐ซ and ๐จ๐ซ๐ญ๐ก๐จ๐ ๐จ๐ง๐๐ฅ ๐ข๐ง๐ฃ๐๐๐ญ๐ข๐จ๐ง: two ways to shape those fixed points
๐งต Long thread ahead (0/n).
Matthew Yang retweeted
๐ฅณ xLLM intro video, made entirely by Claude Opus 5.5.
Best viewed on desktop; on mobile, rotate to landscape.
(Sorry, the voice is robotic; better one next time.)
Matthew Yang retweeted
Wrote a short blog about the "shape" of language models, and the tradeoffs they may present in the future. I genuinely think it's a valuable research direction to start thinking about now, especially w.r.t. harness design.
alexzhang13.github.io/blog/2โฆ
Matthew Yang retweeted
Starting tomorrow, our lab will hold an open-source week: 2 software frameworks, 4 papers, all building a coherent ecosystem.
The theme: Frontier AI on Hardware You Own
Blog post: timdettmers.com/2026/09/21/dโฆ
SoTA results in:
Autocompaction
Autonomous Research
Model compression
Test-time scaling
Deep Research
And a healthcare RL environment giving you a new level of complexity to train healthcare agents.
The ecosystem that we will release is built to be as usable as possible. Agent sessions that run overnight and go on for tens of millions of tokens is made easy. Model compression of a model is automatic: you just get a good model that runs fast locally -- no expertise required. An autonomous research that works out of the box.
Our ecosystem enables a new level of work that can be done locally.
Matthew Yang retweeted
Really fast and great for everyday coding!
Matthew Yang retweeted
This isn't the most notable aspect of today's news, but on the user data issue, there are different kinds of *training on user data* with very different privacy/IP implications. Sadly, AI cos don't like to disclose what they're doing.
- pretrain on user data, with users' tokens as prediction targets: high regurgitation risk, improper
- use user prompts to distill large models into small ones: low regurg. risk, some companies probably do this
- use user traces to construct RL tasks: low regurg. risk, because RL has low memorization abilities, but can extract customer IP, depending on how it's done. Ranges from benign "use explicit user feedback in reward model training" to invasive "upload user's coding environment and commit history to turn into rl envs"
"De-identification" is weak -- you can identify someone with a small number of bits, and long traces have more than enough. And it doesn't affect IP leakage concerns.
