๐ Introducing OpenVLHarness: Turn any LLM into an evidence-based reasoner that thinks and codes in visual space.
๐ GPT-6 Astra gains +9.9 points on average over 22 datasets.
โก Beats Codex and other VL harnesses by 7.7โ8.5 points, at โค37% of their cost.
๐งต More details โ
Weโre excited to officially introduce Noeri!
Noeri enables your company to build specialized AIs from its own data and know-how. The AIs keep learning from real work, team up on your hardest problems, and remain entirely yours.
Founded by @YejinChoinka and @Guestrin, our team brings together researchers and engineers from the most prestigious institutions in the world. Weโre united by one belief: the future of AI should be plural.
Weโre proud to have raised $110M to bring this vision to life, with the support of our investors and partners, including @MenloVentures, @MadronaVentures, @nvidia, @Cisco, @Snowflake, @KindredVentures, @NEA, and @Wing_VC, among others.
And weโre just getting started.
Join us in building the future of plural intelligence!
The End of AI Frontiers https://t.co/rQOIpmthGr .
There's a bit of a train wreck here that's painful to express. Nevertheless, it's a significant event for me and quite a few other folks involved so I wanted to memorialize a bit.
Thrilled to share that our paper DISCO: Distilled Tree Search over Continuous Actions has been selected as a ๐๐จ๐๐ ๐๐๐๐ ๐๐๐ฌ๐ญ ๐๐๐ฉ๐๐ซ ๐๐ฐ๐๐ซ๐ ๐ ๐ข๐ง๐๐ฅ๐ข๐ฌ๐ญ! ๐ Congrats to @_ksaha ๐
Can robots learn to plan for the long horizon without expensive search at deployment?
DISCO uses Monte Carlo Tree Search in simulation and iteratively distills its discoveries into a policy that can execute directly, without search at test time. We demonstrate strong performance on long-horizon object packing and stacking, including real-world transfer!
๐ Catch us in Austin!
๐ค Oral: Nov 10, 11:00โ11:30 AM | Hilton Austin (Session O6)
๐ Poster: Nov 10, 5:00โ6:30 PM | JW Marriott Austin (Session P2, Poster #176)
๐ https://t.co/SXgvbkcQTj
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! ๐งต
Blog: https://t.co/bOWCcFhCVq
Introducing Game2Player ๐ฎ Scaling gameplay experience for visual agents.
Instead of a handful of games or costly human play data, we use coding agents to turn 568 web games into 3,436 verifiable tasks, then collect 1.56M gameplay steps with both code policies and VLM policies.
๐ Environments and players improve together: progress is checked from game state at every step, and any problem a player hits goes back to the builder agent for repair.
๐งช We also introduce PlayerBench: 50 held-out games, 287 tasks. Even GPT-6.1 Sol reaches only 62.6% progress.
๐ Our 9B VLM, trained from our gameplay data, plays from pixels alone. It lifts progress from 30.0 to 53.1 on GameWorld (ahead of Qwen3.8-Flash-Next) and more than doubles it on PlayerBench, from 14.3 to 31.2.
๐ Website: https://t.co/3mjIychogf
๐ป Code: https://t.co/hCRiYRHs2u
๐ค Models: https://t.co/tq0yNkeFrr
๐ More games and results coming soon. Stay tuned!
(๐ต๐๐๐๐๐๐๐ข๐๐ ๐๐ข๐ ๐๐: ๐ ๐๐๐ก๐ก๐๐ ๐๐ข๐๐ก ๐๐๐๐ฆ๐๐ ๐๐ฆ ๐๐ ๐๐๐ ๐กโ๐ ๐๐๐๐๐ก ๐๐ก๐ ๐๐๐ ๐ถ ๐ธ๐๐๐๐ฆ ๐กโ๐ ๐ฃ๐๐๐๐ ๐ค๐๐กโ ๐ ๐๐ข๐๐ ๐๐! ๐)
Ever dreamed of a robot that could play the harp ๐ผ, tie your shoelaces ๐, or even perform delicate wet-lab experiments ๐งช?
The key is ๐ฅ๐๐๐ซ๐ง๐ข๐ง๐ ๐ง๐จ๐ญ ๐ฃ๐ฎ๐ฌ๐ญ ๐ก๐จ๐ฐ ๐ก๐ฎ๐ฆ๐๐ง๐ฌ ๐ฆ๐จ๐ฏ๐, ๐๐ฎ๐ญ ๐ก๐จ๐ฐ ๐ญ๐ก๐๐ฒ ๐๐๐๐ฅ ๐๐ง๐ ๐๐จ๐ง๐ญ๐ซ๐จ๐ฅ ๐๐จ๐ซ๐๐๐ฌ. We introduce ๐ ๐๐๐, a ForceโActionโTrack foundation model that learns from 2,000 hours of force-rich human demonstrations in our EgoForce dataset, bringing us closer to robots with truly human-like dexterity! ๐ฉท๐๐
https://t.co/iR7NQXBELs
Bigggggg shoutout to @ZhiCao01756604 !
Apple and UCLA present OpenVLHarness
A training-free harness that gives multimodal LLMs a capability layer and persistent multimodal memory, boosting average scores by 9.7-12.6 points across 22 benchmarks.
@swordBluesy@omarsar0 Yeah, of course. We have well-maintained code with detailed instructions teaching you how to use the code. You can also work with Codex or Claude Code.
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
๐น 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
๐น Full-project coding: games in Unreal, Unity & Godot.
๐น Excels at parametric 3D modeling in Blender & CAD.
๐น More reliable at finance work: Accurate modeling with hallucination down by 65%.
๐น Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2โ3ร faster) are now live on https://t.co/85jA71X9gw, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
๐ API: https://t.co/XCrgjXAqMw
๐ Tech blog: https://t.co/5iPP0MBej6
Happy to see our paper is listed in Hugging Face trending papers @HuggingPapers https://t.co/98eoxVf6PY
Huggingface: https://t.co/5JaSqpDl7d
Paper: https://t.co/g7S07ZwDkG
Website: https://t.co/Rj5I57NJYx
Code: https://t.co/zbNWFiOD5w
Demo: https://t.co/ZoF5FdJEyU
๐ Introducing OpenVLHarness: Turn any LLM into an evidence-based reasoner that thinks and codes in visual space.
๐ GPT-6 Astra gains +9.9 points on average over 22 datasets.
โก Beats Codex and other VL harnesses by 7.7โ8.5 points, at โค37% of their cost.
๐งต More details โ
I'll be at EMNLP next week with Liu (1st-auth) to present this work (main conf + WMT)
I'll also be presenting the short paper "Large Reasoning Models Struggle to Transfer Parametric Knowledge Across Scripts" w/ Trevor Cohn https://t.co/X5yEPMtIvX (findings)
DM me to meet !
๐งตDoes a correct answer imply faithful compression?
Introducing ๐KEEP-Bench๐: a benchmark that directly tests whether context compression preserves all the evidence needed for reasoning, exposing unfaithful compressions that traditional QA evaluation misses.
1/n
Thanks AK @_akhaliq for featuring our work!
Excited to share OpenVLHarness, a multimodal agent that inspect, compute, and reason over evidence.
The technical core:
โข A capability layer gives consistent interfaces for various tools such as perception and search, and renders their outputs into observations the model can inspect.
โข Persistent multimodal memory retains images, masks, numerical arrays, and web references under reusable handles.
โข A coding agent computes over those artifacts to combine evidence across tools.
Across 22 benchmarks and 8 backbones, we see average gains of 9.7โ12.6 points over matched no-tool baselines. Evidence acquisition and additional reasoning are complementary!
A joy collaborating with this UCLA ร Apple team. Huge shoutout to @yu_bryan_zhou, @cfyang58, @RuiSun94013021, @DiWu0162 , Sibo Peng, Mingyang Zhang, Zhen Yang, Ruiyi Zhang, @zhegan4 and @kaiwei_chang! ๐