Pinned
Ai2
3,744 posts
Ai2
@allen_ai
Breakthrough AI to solve the world's biggest problems.
› Join us: allenai.org/careers
› Newsletter: tinyurl.com/3vc2r2m8
- Ai2 reposted
New work from @YanhongLi2062 digging into the specific tokens that hybrid models predict better than transformers 📈 Spoiler alert: gains are broad across token categories, especially large on content words. Gains diminish on copying tokens, but even there hybrids aren't worse 
We hope this kind of token-level comparison helps the field build better, more architecturally diverse models. Going forward, we plan to explore applications of these findings to pretraining evals. ✍️ Blog: allenai.org/blog/hybrid-to… 📄 Tech report: arxiv.org/abs/2606.20936- MolmoSpaces is our open leaderboard for robot policies. Congrats to @NVIDIARobotics on the current top spot. Anyone can submit a policy and try to beat it: molmospaces.allen.ai/leaderboardCosmos 3 wins another leaderboard. The power of an open frontier model. 🎉 Ranked #1 on @allen_ai's MolmoSpaces robot policies leaderboard, a policy model post-trained on Cosmos 3 Nano shows the strongest performance in environments that require models to interpret, reason, and
- Replying to @allen_ai @aisquaredai and @domynaiIn AISquared's platform, Bolt now handles request routing & acts as a policy guardrail. For Domyn, our Dolci dataset added 10.1 points to Domyn-Small on GPQA-Diamond, a graduate-level science reasoning benchmark—its biggest single post-training gain.
The throughline: for AI labs working in high-stakes domains, the bar isn't just capability—it's transparency & control. Ai2's full openness is what puts that within reach. 📝 Read more: - Ai2 reposted
MolmoMotion is not just another cool work built on top of Molmo2’s video grounding capabilities. It shows grounding in 3D can enable motion forecasting (for both data and capabilities), which serves as an objective and representation for controllable video generation and
00:00 - Ai2 reposted
Thrilled to introduce MolmoMotion🚀where we found 3D motion prediction is becoming very doable with existing tools. And this capability can be very useful for downstream tasks like Robotics📷and VideoGen📷 As always, MolmoMotion is fully open! Great work led by @jianingnz00:00 - Ai2 reposted
Very cool project with nice applications of motion prediction in video generation and robotics! Also cool to see positive transfer from Molmo2’s video pointing & tracking capabilities to motion prediction 📈00:00 - Ai2 reposted
We released MolmoMotion today, and I’m excited to finally share it! Give it a frame, mark a few points on an object, and say what you want to happen. It traces where they travel in 3D over the next few seconds. The weights, data, and benchmark are all open, so give it a try 👇00:00 - Ai2 reposted
Another strong proof point for upgrading VLMs into 3D foundation models, and ultimately supporting robotics. Huge shout-out and congrats to the team!00:00 - Ai2 reposted
Language gave foundation models semantics. Vision gave them grounding. Motion may be the next ingredient. Molmo Motion explores learning from the dynamics of the world itself capturing how objects, people, and agents move through time. Check out our latest work MolmoMotion!00:00
