Siqiao Huang
288 posts
@KnightNemo_
Rising senior undergrad, Yao class @Tsinghua_Uni | Current intern @uwcse | World Models, WAMs, Humanoid Foundation Models | RS @wuji_global | Prev. RA @mldcmu.
Everyone is talking about self-evolving AI, or recursive self-improvement. The methods that built modern ML are the ones that keep working across settings and scales, yet no benchmark directly tests AI systems for that ability. Today, after months of cross-platform validation with external teams, we're proud to introduce MLS-Bench, a benchmark for ML Science with 140 tasks across 12 ML domains. It asks whether AI systems can create scalable and generalizable ML methods, the way human researchers have pushed AI forward. (1/9)
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)
🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over K2.6: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite. 🔷 Reasoning efficiency: Less overthinking, with 30% lower reasoning-token usage compared to K2.6. 🔷 Long-horizon coding: Improved instruction following, higher end-to-end coding task success rates. ⚡️ 6x High-Speed Mode coming soon! 🔌 Available today via Kimi API and Kimi Code. 🔗 Kimi Code: kimi.com/code 🔗 API: platform.moonshot.ai
The GPT-4o moment for humanoids might finally be here. And yeah, sorry in advance for the rickroll. OMG runs a Unitree G1 off one brain that natively takes language, audio, and human motion. That "Never Gonna Give You Up" dance isn't the flex; one model fluent in every modality is. Here's the shift. Most humanoid policies are one-trick. Train per skill, hand-tune the rewards, repeat. The rest just replay a fixed motion you feed them. OMG instead works like a biological motor system. A "brain" that turns intent into future motion. A "cerebellum" that reactively runs it on the robot. The brain is one diffusion model. Language, audio, a reference pose, or any blend goes in. A robot-ready G1 trajectory comes out, live. New inputs attach through zero-init adapters. They start at zero, so the pretrained motion prior carries over intact instead of getting scrambled. That's how they bolted on VR teleoperation as a brand-new modality, reusing the same brain. And it behaves like a foundation model. Bigger backbone, cleaner motion. Finetune on 1% of new data, nearly match a model trained from scratch on 100%. Compose language + audio at inference for combos never seen in training. 1000+ hours of motion, all retargeted into one G1 body. One brain that scales. We keep racing to build stronger low-level controllers. OMG's bet is that the real bottleneck is the brain mapping human intent to motion. Congrats to @KnightNemo_, @li_yitang, @ShaotingZ38103 and whole team!
