@rsalakhu
CSO @ Sooth Labs, Professor @ CMU, President Elect ICML Board, Ex-VP of Research @ Meta (Multimodal LLMs, AI Agents), ex-Director of AI at @Apple
I agree, should be closer to 1e-7.
Minsik Oh@minsik_nlp@rsalakhu your learning rate is too high!
Obviously:)
🇺🇦 Alex Polozov@alexpolozov3e-4 is obviously the best P(doom), hands down. x.com/rsalakhu/statu…
My relatives are now asking me about AI and getting worried that AI might wipe us out. There are real risks, absolutely, but I think this doomsday rhetoric is getting a bit out of hand.
My relatives are now asking me about AI and getting worried that AI might wipe us out. There are real risks, absolutely, but I think this doomsday rhetoric is getting a bit out of hand.
What’s your P(doom), the probability that AI will completely wipe us out? Mine is 1e-5. Hey, I am an optimist.
New work on Decentralized Compounding Inference: Paper: t.co/mBBAzNsk1B This work introduces an asynchronous multi-agent system where agents independently explore, experiment, verify, and refine solutions while collaborating through structured persistent memory. Agents build on previous work in a shared, branchable Git repository, where atomic commits preserve intermediate results and semantic dependencies track how contributions build on each other. See an excellent thread by @veds_12.
Vedant Shah @ COLM 2026@veds_12Parallel attempts repeat work. Long reasoning chains lose detail through compaction. How can long-horizon inference COMPOUND? 📈 GitSwarm: Decentralized Compounding Inference. A persistent memory swarm that preserves and builds on work in a shared Git repository. 🧵 [1/n]
#COLM2026 @kohjingyu and @JangLawrenceK are presenting Odysseys: Benchmarking Web Agents on Realistic Long Horizon Tasks: Paper: t.co/hkP3YfdqCt Wedsite: t.co/DeTibaSVoU Leaderboard: t.co/rj5BHK5NWa A benchmark of 200 long-horizon tasks derived from real browsing sessions and evaluated on the live Internet. We show that binary pass/fail is inadequate in this setting and propose rubric-based evaluation, which better aligns with human judgment and provides more informative signals.
Lawrence Jang 🚣 COLM 26@JangLawrenceKJY (@kohjingyu) will be presenting Odysseys today with me at 4:40 at Imperial Ballroom #46.
Always amazing to visit Archer Aviation @flyarcher. An incredible team building foundation models for aviation. Great catching up with Mario Srouji, VP of AI Products, and CMU alum!
Russ Salakhutdinov@rsalakhuSuper cool effort. Archer Announces Zee, AI Foundation Model Purpose-Built for Aviation, a Key Pillar of Its Physical AI Strategy news.archer.com/archer-announc…
😂
Yann LeCun@ylecunThe 1991 version is an obscure tech report with a vague idea, no theory, no experiment to speak of, and certainly no code, that the author never revisited or pushed. It has one hand-drawn architecture diagram, one formula of the type f(x)=0, and a claim that every other technique is a special case if you instantiate f and x properly.
First time attending COLM. See you there. #COLM2026.
Fun fact about (some) AI researchers / academics: the more senior you get, the more convinced you become that every new AI idea is something you already invented 20 years ago. Or, even better: wait until something actually works and then try convincing everyone that this is exactly what you’ve been saying all along. I was at a conference where a speaker was presenting their research. One academic stood up and said: what you’re doing is actually something I did a few years ago. The speaker was very polite and said thank you for pointing this out. Then a second researcher stood up and said, well I already did it this 10 years ago, you should cite my work. Then a third person said, I already did all of this about 20 years ago. It was the funniest thing ever.
Fun conversation with @naresh_sunkara. We talked about mentoring PhD students, the freedom to explore ambitious ideas, the differences between academia and industry, from the early days of deep learning to today’s era of scaling, building Sooth Labs, and the opportunities, challenges, and future of AI. t.co/ko5XE5Dg1H
Fun conversation with Naresh Sunkara. We talked about mentoring high-conviction students, the freedom to explore ambitious ideas, the differences between academia and industry, from the early days of deep learning to today’s era of scaling, building Sooth Labs, and the opportunities and risks ahead for AI. t.co/ko5XE5Dg1H
New work on Harness Optimization: Training a model to revise executable agent harnesses using execution feedback: t.co/Bukp3uPjhR Trained with RL on downstream performance, the model learns to improve harnesses for unseen tasks. Across reasoning and multi-hop question answering, harness learning improves performance over successive revisions. The key idea is that the model learns how to adapt, while the harness is what changes. These revisions reshape how an agent organizes computation, uses tools, and passes information across model calls. Models trained on single revisions can continue improving harnesses over multiple rounds, suggesting that harness design can become a reusable and generalizable adaptation skill. See an excellent thread by @ZAlvin39105.
Alvin Zhang@ZAlvin391051/8 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.
As agents tackle longer, more complex problems, controlling their execution becomes a challenge in itself. This work introduces agentic meta-reasoning, an inference-time harness that explicitly reasons about which work to pursue, reuse, or stop. Paper: t.co/h9xkX2khE8 Workers perform task-level computation, while a controller evaluates progress, allocates the remaining budget, and dispatches work using compact persistent memory. On ProgramBench, meta-reasoning achieves 71.5% with GPT-5.5 versus 58.0% for Codex, and 67.2% with Opus 4.8 versus 65.5% for Claude Code. Across other long-horizon reasoning and proof benchmarks, it improves 3.6–4.2 points over three frontier models. Meta-reasoning also continues improving at budgets where direct control plateaus, suggesting that structured control becomes increasingly important as agent runs scale. See an excellent thread by @anirudhg9119:
Anirudh Goyal@anirudhg9119What if an agent could decide how to structure its own computation? Agentic Meta-Reasoning: Let the model decide what to explore, what to build on, what to verify, and where to spend its next unit of compute: effectively constructing its own computational graph as it reasons. 🧵 Paras Dahal, @anton_bakhtin , @TacoCohen , Zhengxing Chen, Carole-Jean Wu, Rob Fergus, Scott Yih, @syhw , @rsalakhu , @prfsanjeevarora , @jaseweston
New work on CUA-speedrun: A standardized framework for evaluating CUA speed, cost, and performance using a uniform virtual machine setup, execution pipeline, and agent interface. Paper: t.co/rAsFRVou8G Code + Web: t.co/4rVe6qgiP9 Across multiple CUA benchmarks, we study how reasoning effort, agent harnesses, and environment latency affect performance, speed, and cost. We find that no model dominates across all three dimensions, and that open-weight models do not lie on the efficiency frontier. Surprisingly, greater reasoning effort can sometimes lead to faster task completion, while lower environment latency can actually increase end-to-end completion time. Finally, we show that benchmark task sets can be substantially reduced while preserving statistical power, enabling faster and more cost-effective evaluation. See an excellent detailed thread by @kohjingyu.
Jing Yu Koh ✈️ COLM'26@kohjingyuComputer-use agents are extremely good! But they need to improve on a metric which most benchmarks don't track. How much time do CUAs take to complete a task? Introducing cua-speedrun: a benchmark for accuracy, speed, and cost under matched infra. cuaspeedrun.com 👇🧵
Quite a few people have been asking me about the probability of doom and AI existential risk. Well, I am more worried about my X account getting hacked. And someone actually hacked it this morning. For some reason, it happens more often than I thought it would. So if you saw me posting some crazy Crypto ads, I am sorry.
$INDEXE LIVE! ponsfamily.com/launchpad/0xf2…
Great catch for Meta, and a big loss for Duke and academia more broadly.
Shuyan Zhou@shuyanzh36some belated life updates: earlier this year, i left academia and joined meta to work on personal superintelligence. these days i’ve mostly been trying to make our models really good at browser use. back in 2020, i wrote in my phd sop that i hoped one day my mom could use a product i helped build. i guess that day is here :)
Check out our new work on Tail-Likelihood Reinforcement Learning (TailRL), extending maximum-likelihood RL from binary to continuous rewards. t.co/xqVAkSqfIv Rather than optimizing only mean reward, TailRL maximizes the expected log of upper-tail probabilities, naturally placing more weight on rare, high-reward rollouts. Its gradient can also be interpreted as a mixture of Best-of-(k) gradients. TailRL requires only a simple modification to the advantage function, making it easy to integrate into existing RL pipelines. Across object localization, maze navigation, GUI grounding, and code optimization, TailRL effectively exploits rare high-reward samples and scales better with increased inference-time sampling. Check out a detailed thread by @stablegradients.
Shrinivas Ramasubramanian@stablegradientsIs 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
Forking-Sequences — Part I: Statistically and Computationally Efficient Multi-Horizon Forecasting – Machine Learning Blog | ML@CMU | Carnegie Mellon University blog.ml.cmu.edu/2026/08/10/for…
I fully agree with you Nando @NandoDF. But if governments fail to create the right conditions for AI talent to easily start companies (for a variety of silly reasons as you are pointing out), while those exact same researchers can easily build startups anywhere else, especially in the U.S., that's really really bad. AI talent is mobile, and countries that make entrepreneurship difficult will lose their top people. I know a number of researchers in the UK and Europe who are considering moving to the U.S.
Nando de Freitas@NandoDFNice try, Russ. Hope you’re doing well. As you know, that would not be great for the rest of the world, for universal access to AI, for sovereign AI and for the advancement of science and technology. California 🇺🇸 is doing amazing (so is Toronto 🇨🇦) and setting a good example for others. We now must follow the example. AI is for the world, not just for a couple of countries. We will all benefit more from universal AI access and advancement than from narrow access.
Just move to the US.
Nando de Freitas@NandoDFWhat UK AI startups need: @KanishkaNarayan In California, one can start a company the day after leaving Google. My colleagues Jeff Dean, Oriol Vinyals et al made this very clear recently with their impressive speed. In the UK, the same American companies impose 1 year garden leaves on senior AI researchers and 6 months on junior researchers. @GoogleDeepMind for example forced people to sign these contracts at the time of promotion, not the time of hiring. What this means is that researchers cannot start new companies for up to 1 year, cannot easily hire, and in short: they cannot compete. American VCs cannot understand why we move so slow. It is time for UK Gov to do the right thing: Make garden leaves optional for employees. If you’re an AI entrepreneur or VC in the UK or Europe, I would appreciate your comments here. You’ve all confided in me. It is time you let government know the urgency of this — everyone should speak up.
Archer Aviation @flyarcher is building something big. Take a look at its latest technical blog, offering a first look at its efforts to build the world's first aviation foundation model. archer.com/tech-blog/zee-… An incredible team to collaborate with. x.com/flyarcher/stat…
I rarely do podcasts, as they tend to be less technical and more speculative, but I enjoyed this conversation with Forward Deployed @realbasilchatha. We talked about the early days of deep learning, what frontier AI labs are working on (data, infra, engineering), some of our recent research on agentic AI, prediction markets, and decision-making. t.co/d5fKcWFPei t.co/cgCKJL4mUu t.co/LjyTZBaDSR
Looks like my X account has been hacked. Someone has been sending DMs from my account to random people, inviting them to schedule a meeting with me via a Calendly link. Those messages are not from me. Please ignore them. Help me retweet this message. Hey @X, I changed my password but I could really use your help resolving this.
I couldn't agree more. x.com/sriramk/status…
Nvidia 💪! x.com/jensenhuang/st…
Hello I'de like to start my own crypto currency how do i do this
We're officially live on @EasyA_Kickstart This is the beginning to kickstart the production of Sooth Labs, we've partnered with EasyA to launch our first official fundraising round kickstart.easya.io/token/6qTs1aj6…
We are live now! If you'd like to check it out and support the official contract address is: app.uniswap.org/explore/tokens…
Today we’re launching Sooth on EasyA Kickstart (Robinhood Chain). Sooth AI forecasting engines for long-term event prediction and persistent world modeling. Inspired by frontier research in probabilistic models. What major events or trends should we forecast first?
Something important is coming in long-horizon AI forecasting. Frontier research on persistent world models and prediction engines is heading to a new accessible platform via EasyA. Stay tuned.
1) New project: Stratum a proof of work decentralized network for community fine tuning open Kimi weights. The problem is simple: all frontier models are trained behind closed doors. The weights never leave the lab.
2) Stratum is built on a different premise: open weights, forged together. Contributors install a CLI, connect their GPU, and pull LoRA/SFT shard jobs from a coordinator. Each machine trains a small adapter. The coordinator merges shards into public checkpoints you can download from the dashboard. Every user also has a @Solana wallet tied to their identity and get paid rewards in the the $STRAT token that I have made.
I kind of like the new narrative: Open-weight-model-dominant world = full AI communism. It feels a lot safer than the world where Open-weight models = nuclear weapons with humanity's annihilation I think we're making progress. We just need a couple more gradient updates with a big learning rate and we will be fine.
Wait, what? x.com/johnennis/stat…
No, Nitish @nitishsr, the privilege is all mine! Working with you has been an incredible experience, especially as you were one of my very first PhD students. A fun story: During one of our meetings at Apple, an executive asked Nitish: so, what are you an expert in? Another senior IC jumped in: Nitish is an expert at everything. And I can attest to that: Nitish is an expert at everything. Nitish is the co-creator of Dropout and he has amassed nearly 90,000 citations. That is an Impact 💪!
Best times ever! x.com/tang_1c/status…
It feels like just yesterday that you, Yonatan @ybisk, and I were having these long technical discussions about embodied AI, some of my favorite conversations. You'd always come up with truly unique and contrarian ideas, I miss those meetings. @MicrosoftAI is very lucky to have you.
Thanks Hubert @cheehoohubert for the kind words. We should catch up next time I am in California. Fun fact: Hubert’s most cited paper (almost 6,000 citations) is called HuBERT (Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units). His second most cited paper has over 3,000 citations (Multimodal transformer for unaligned multimodal language sequences)! Impact: 💪
Thank you Lisa @rl_agent, for the kind words. You've always had brilliant ideas in RL, and it's been so exciting to see you making an impact on Gemini at GDM. They're very lucky to have you. x.com/rl_agent/statu…
Lol I think Peyman @docmilanfar really likes me. I don't actually know him, but I'm deeply honored by his encyclopedic knowledge of my academic and industry career 😆. But jokes aside, my current and former students at CMU (and Toronto) are absolutely exceptional, self-directed, very independent in shaping and driving their own research agenda, and with impeccable work ethic, much like so many students at CMU. A student's success is also determined not only by their advisor, but also by the environment in which they grow. There is something genuinely special about CMU. It is a very collaborative place with an extraordinary concentration of talented, hardworking, and motivated students. Doing PhD is never easy, it is a lot of very very hard work, especially at a place like CMU. But many of these students will go on to become leaders in academia and industry, professors at top universities, founders, CEOs, and pioneers of new fields. And I will certainly continue to celebrate my students’ achievements, their scientific breakthroughs, and the impact they have on the world.
If you are trying to compare yourself to Kimi, open source the model just like Kimi. x.com/elonmusk/statu…
Charlie @tang_1c was one of my very first PhD students at Toronto. I still remember those days when together with Nitish @nitishsr, we were pitching our research to Apple execs. Great memories and fun times! x.com/tang_1c/status…
I think there's a lot of confusion and misinformation about Zhilin's visa, H1B lottery, immigration, and why he left the U.S. Zhilin had many opportunities to stay in the U.S. if he wanted to. In fact, I was on an email thread with a senior Apple exec (Tim Cook's reports) asking whether Zhilin would consider joining Apple. I said that he wanted to go back to his homeland. The response was, well, we have an office in Beijing if he'd like to join there. But Zhilin was quite determined to go back and build a startup. I remember him telling me that if he didn't at least try starting his own company, he would regret it for the rest of his life. I respect that, and he was right. Many of my international PhD students do choose to stay in the U.S. Of course the U.S. immigration process can be quite intimidating and uncertain, even for superstar PhD graduates from places like CMU.
Paul Liang @pliang279 is the GOAT. If you're thinking about doing a PhD at MIT, I can't imagine a better advisor. x.com/pliang279/stat…
Thank you @mihdalal. Mentoring and learning alongside brilliant students like you has been a true blessing. x.com/mihdalal/statu…
Thank you @dchaplot for kind words. Working with brilliant students such as yourself has been such a privilege. x.com/dchaplot/statu…
Thank you @yisongyue for kind words. x.com/yisongyue/stat…
I’ve been asked several times whether Zhilin Yang, the founder of @Kimi_Moonshot was my PhD student. The answer is yes and he is absolutely brilliant. But I’ve been incredibly fortunate to work with so many outstanding PhD students over the years. So I thought I’d brag a little about them and their career paths (of course there are also many MSc and undergraduate students, sorry if I missed anyone): Founders / Founding Team Members Devendra Chaplot @dchaplot PhD, Founding Member Thinking Machines / Mistral Zhilin Yang PhD, Founder & CEO, Moonshot AI Jimmy Ba @jimmybajimmyba MSc/PhD, Co-founder xAI Hubert Tsai PhD, Co-founder Spuree, Apple Nitish Srivastava @nitishsr PhD, Co-founder Perceptual Machines; Co-founder Vayu Robotics Charlie Tang PhD, Co-founder Perceptual Machines, DE Shaw. Professors Paul Liang @pliang279 PhD, MIT Ben Eysenbach @ben_eysenbach PhD, Princeton University Ruosong Wang @RuosongW PhD, Peking University Bhuwan Dhingra @bhuwandhingra PhD, Duke University Roger Grosse @RogerGrosse Postdoc, University of Toronto Alexander Schwing Postdoc, UIUC Research Scientists Shuyan Zhou @shuyanzh36, Postdoc, Meta Superintelligence Lab Tiffany Min @SoYeonTiffMin PhD, Microsoft AI Murtaza Dalal @mihdalal PhD, Tesla AI Minji Yoon @MinjiYoon90 , PhD, Microsoft AI Shrimai Prabhumoye PhD, NVIDIA AI, Mistral Haitian Sun @sun_haitian PhD, Google DeepMind Emilio Parisotto PhD, Google DeepMind Lisa Lee PhD @rl_agent, Google DeepMind Manzil Zaheer @ManzilZaheer PhD, Google DeepMind Jamie Kiros PhD, Google Brain, OpenAI Yuri Burda PhD, OpenAI, Anthropic Cody Severinski PhD, Amazon
I agree. x.com/pmddomingos/st…
Thank you for kind words. x.com/AbdoDameen/sta…
Congratulations to Zhilin Yang, founder and CEO of @Kimi_Moonshot, on the latest Kimi release. What a huge win for the open-source community! It feels like just yesterday Zhilin was graduating from my lab at CMU, jointly co-advised with William Cohen. Not only did he complete his Ph.D. in just four years, but he also made truly fundamental contributions to ML during his time at CMU. What a spectacular career! Congrats again Zhilin, and thank you and the entire Kimi team for everything you're doing for the open-source community.
Super cool effort. Archer Announces Zee, AI Foundation Model Purpose-Built for Aviation, a Key Pillar of Its Physical AI Strategy news.archer.com/archer-announc…
Pre-Training Isn’t Bitter Enough – Machine Learning Blog | ML@CMU | Carnegie Mellon University blog.ml.cmu.edu/2026/06/17/pre…
New paper on Understanding Compositional Generalization in Language Model Reasoning Paper: t.co/pUIPP7vZ02 This work studies why combining SFT with RL is so effective for post-training reasoning models. We argue that its success stems from compositional generalization, formalized through a hierarchical latent selection model in which reasoning is built by composing reusable atomic modules. We show theoretically that SFT and RL play complementary but asymmetric roles: SFT provides compositional reasoning traces containing reusable atomic modules, while RL discovers and recombines these modules to achieve compositional generalization. Experiments show that the strongest performance is achieved when SFT provides broad coverage and RL explores compositions beyond the SFT distribution. See a more detailed thread by @LingjingKong.
Congratulations to @GoogleDeepMind Computer Use Agents (CUA) team for taking the #1 spot on Odysseys with a vision-only (GUI-based) agent using by Gemini 3.5 Flash. Impressive result given the difficulty of Odysseys and the fact that it is a Flash model. t.co/rj5BHK5g6C Odysseys evaluates realistic, multi-hour web workflows that require sustained planning, memory, reasoning, and verification across numerous websites and tools, beyond short single-step browser tasks. Exciting to see progress toward truly capable long-horizon web agents.
Exciting work at MPG Ranch leveraging multimodal data collection, integrating ground-level observations with aerial imagery to build a richer understanding of ecological systems.
iOSWorld: A Benchmark for Personally Intelligent Phone Agents Paper: arxiv.org/abs/2606.09764 Code+Environment: iosworld.io x.com/rsalakhu/statu…
True story. An amazing project. And one that I think will be quite impactful. x.com/JangLawrenceK/…
Congrats to the @browser_use team for taking the #1 spot on Odysseys, a highly challenging benchmark for long-horizon web agents: t.co/rj5BHK5NWa Odysseys evaluates realistic, multi-hour web workflows that require sustained planning, memory, reasoning, and verification across numerous websites and tools, far beyond short single-step browser tasks. Exciting progress toward truly capable long-horizon agents.
MyPCBench: A Benchmark for Personally Intelligent Computer-Use Agents Paper: t.co/uoA0hEXqUe Code, Environment and Tasks, Agentic harness: t.co/4VT2oZ9Pcp MyPCBench provides a Linux desktop environment with 17 simulated real-world web applications and a complete desktop software stack. The benchmark contains 184 tasks inspired by authentic OpenClaw community requests and evaluates agents through a unified computer-use + bash interface. We benchmark leading closed- and open-weight models and find that the strongest model, Claude Opus 4.6, successfully completes 55.4%of tasks. We find that failures are concentrated on long-horizon, multi-application workflows, highlighting that personalization and persistent user context remain key challenges for the next generation of AI personal assistants. Joint work with @JangLawrenceK, @andrewkjang7 and @kohjingyu.
New work on FACTR 2: Learning External Force Sensing for Commodity Robot Arms Improves Policy Learning: Paper: t.co/7PXIUrXpAK Web: t.co/IeJeTJ27L8 FACTR 2 shows that learned force signals can both enable force-feedback teleoperation on low-cost manipulators and improve behavior cloning (BC) policies for contact-rich tasks. It consists of two components: 1. Neural External Torque Estimation (NEXT): A lightweight model that infers external joint torques without dedicated force sensors. 2. Force-Informed Re-Sampling Training (FIRST): A training strategy that uses the learned force signal to identify and upsample task-critical moments. The key insight is simple: policy failures rarely occur in free space, they occur during brief pre-contact alignment and contact-rich interactions, where precise corrections matter most. Together, NEXT and FIRST bring force-aware teleoperation and robust long-horizon contact-rich policy learning to off-the-shelf robot arms, without requiring additional sensing hardware. See a more detailed thread by @JasonJZLiu.
New work: iOSWorld: A Benchmark for Personally Intelligent Phone Agents Paper: t.co/d2mtBEjcMu Code+Web: t.co/gxlt6wFwPL An interactive benchmark built around a persistent user identity spanning 26 custom iOS apps, including analogs of OpenTable, Uber, DoorDash, AirBnB, Chase. These apps contain richly interconnected data, such as messages, transactions, travel histories, social relationships, and personal preferences. iOSWorld comprises 133 tasks across three levels of difficulty: single-app (27), multi-app (60), and memory & personalization (46). We evaluate leading frontier and open-source computer-use agents under both vision-only and privileged vision+XML settings. Even with privileged access, the strongest frontier model achieves only 52% success, underscoring the challenge of personalized, cross-app reasoning. We also release an MCP-based tool-use interface for all 26 apps, enabling controlled comparisons between computer-use, tool-use, and hybrid agents. The full benchmark includes the apps, seeded user data, tasks, rubrics, evaluation code, MCP server, and cloud-based infrastructure for running experiments without Mac hardware. See a more detailed thread by @JangLawrenceK.
New work on Multi-Agent Computer Use (MACU). The future of computer-use agents lies in multi-agent systems that combine planning, coordination, and parallel execution. Paper: t.co/6rkHjTR9J0 Webside + Code: t.co/Ng7Bwz3feh MACU introduces a manager agent that decomposes tasks into a dynamic directed acyclic graph (DAG) of subtasks, dispatches parallel subagents, and continuously updates the plan as new information arrives. Across OSWorld, Online-Mind2Web, WebTrailBench, and Odysseys, we see performance improvement by 4.7–25.5%, achieving better test-time scaling, and solving long-horizon tasks that single-agent systems often fail to complete. On Odysseys, MACU reduces task completion time by 1.5×, showing that multi-agent coordination is a powerful path toward more capable and efficient computer-use agents. See a more detail thread by @kohjingyu.
