Stanford's AI is Self-Learning w/ Context Engineering: ACE

hello community so great that you are back let's talk about agentic context engineering and yes welcome to the channel discover.ai we have some brand new research now if you see an x here there is a new study here by stanford university uc berkeley and they say end of fine tuning another game-changing breakthrough rest in peace fine tuning Stanford just killed it here. Everyone obsessed with short clean prompts. But now we go for long detailed playbooks that never forget. So you ask me, hey, what is it? What is this new study by Stanford University? Is it true that fine tuning is gone? Well, not so fast, my friends, because of course it is a little bit more complex. So let's start the video. We have our classical methods here of post-training like supervised fine-tuning and expert data that limited the generalization due to the data scarcity and the lack of exposure to failures, while our reinforcement learning methodology are often infeasible in reward-sparse environments. Therefore, there is now this new trend that we say, hey, is there something else except supervised fine-tuning and reinforcement learning for a particular data configuration? And yeah, if you combine two AI research papers, I think there's no indication that something new is approaching us here in artificial intelligence. So we are looking for low overhead alternatives that leverage here the model's own outputs and the model's interaction. Why? If I don't have the training data of supervised fine-tuning, and I do not have the alignment data of the reinforcement learning by verifiable reward structure, so what do I have on data? I have that the model must generate some internal created data. So the question is now how we do this. And now guess what? That makes no sense if you just put AI into isolation and then the AI reflects on its internal knowledge. Nothing new is going to happen. But if you put AI in a real world environment and you let AI interact massively with this environment, the AI will learn on real world application. This is not a particularly beautiful if we have environments that change. If we have environments that change real fast, environments that evolve or environments where we have no ground truth at all for our reward functionality. So for this case, let's have a look at this. So here we have the paper, Stanford University, October 6th, Samba Nova Systems, UC Berkeley, a beautiful paper about agentic context engineering. Evolving context. from an EI for a self-improving language model. And whenever you see self-improving, you know, oh, wow, this is going to be interesting. But since we are talking about context, I would recommend here, my viewer, keep the context. And the context I would recommend for you is this here. Three days later, October 9, we have here Meta Superintelligence with FAIR. Oh, the first time I see here, Superintelligence and FAIR, who is the lead? the latest superintelligence, okay, Ohio State University, and they have agent learning via an early experience. So if you look at these two papers here, you see we have context of those papers. This early experience I will call the EE paper, early experience, and here ACE is, of course, Agenda Context Engineering. And you know what? This is here, more or less about supervised fine-tuning, that is the perfect optimized input for ACE. So, is supervised fine-tuning that? No, it is just, in my interpretation, the perfect input generator to make this A-system even more powerful. So, let's have a look. Collectively, they advance into self-improving LLM-based AI systems in a reward-sparse environment by introducing a reward-free paradigm. Finally, you might see no output reward, no stepwise reward function. reward free that leverages here an execution feedback and a self-generated data structure and now exactly bridging here supervised fine-tuning and reinforcement learning there is something new combined in the middle above it let's have a look since the technical terms in those papers are not one-on-wallet mapping i don't want to jump here into a technical term discussion i just want to give you an example I want to show you how easy it is if you understand the core principle. So let's start here with a simple example where we integrate both paper. So we are looking now for an AI agent configuration where we have an agent that is now in contact with the real world. So we have a real world experience loop, but directly interacting with any environment, even a simulated one, never mind, understand now the dynamic of this without needing a predefined reward function. This is the beauty. But we need something, no? Let me show you. So since we have two papers, I would say I go above two papers. I go beyond those two papers. Because this should be a funny video, no? This should be a little bit challenging. So let's say we have a two-tier learning system. Level one is E, no? Generates a rich stream of low-level of tactical learning signals about immediate consequences when the AI touches the environment. And then ACE, level two, is about high-level strategic knowledge-based structures. Let's call it a playbook. for general principles and a robust strategy. So you see, those two get beautifully together. Of course, we can build now a cycle. Strategic wisdom from level two makes the tactical exploration in level one much more intelligent and efficient over time. And better exploration leads to better data structure, which lead again to a better playbook. So you see, self-reinforcing beauty. Let's start. We have now... An agent. And an agent, we have to give it, if it has contact with an environment, we have to provide the agent with some knowledge. So we have to supervise, fine-tune it on a small set of expert demonstrations. Let's say 10 trajectories, 10 human trajectories, showing here how the system, how it can do some booking here, a flight booking. So those are the expert trajectory. They might look like this. State zero. It's the homepage with a search form. Then we have an action. Enter the purchase, San Francisco, arrival, JFK, date. Then we have a new state. You see the results page with the flight options. Then we have another action, select the cheapest non-stop flight. Then we have another state, payment page comes up. We have another action, enter your credit card, submit it. And we have another state, confirmation page. Success. That's all we need in an expert trajectory here. to start a supervised fine-tuning to make this AI familiar with the environment it is currently approaching. Let's say we achieve about 60% success for in-distribution tasks. If this is always the same page and we always go with the same parameters, but let's say we go out of distribution, it will drop to 40% because we will have a limited exposure to failures or even edge cases. And now comes here the first paper. Bye. super intelligence. They tell us, hey, we have your two main elements that we discovered. At first, we have to build an implicit world modeling strategy, IWM. And then we do have a critique system where we have a self-reflection of the AI system on its own generated insights. You know this. This is nothing special. This is just generating and reflecting on a self-proposed trajectory set. So we do it. Implicit world model strategy, IWM, you just generate a trajectory. So the supervised fine-tuned initialized agent rolls out now 50 partial trajectories. Because we want to manage longer horizons than we need in the second paper, so the first action by our AI is now. Enter departure, San Francisco arrival, JFK, date, beautiful. Future state result page is valid. We did the booking. Great. Another action could be select a random flight. Then the future state is the payment page was sold out, therefore we cannot do it, we receive an error. Now, on all those actions, 20, 50, 100 actions, the eye just is in contact with the environment and tries it out and gets a feedback from the environment. The agent fine-tunes now on those pairs that it experiences. So what we have, we have a triplet, we have a current state, then the eye takes an action and tries to predict the future state. Then we... Get an error, yes or no. So the AI is learning here the environmental dynamics implicitly, and this grounds here its policy. It predicts that random selection often leads to errors, improving the success to 70% by anticipating a transition like an action, checking the availability first. So you see, the AI is learning by interacting with the environment. The implicit world modeling is rather simple. Our description of the agent fine-tuning is the current state and action and the predicted future state. It's spot on what we need to learn. Now, by learning that selecting a random flight here in the real-world experiment often leads to sold-out error state, the agent is internalizing this dynamic by this environment. But this is exactly the goal here of IWM, to build a grounded policy that has a better feeling of the eye of the world's cause and effect structure. before it even attempts to optimize now for the final goal. And let's say with this optimization and with this continuous fine-tuning, supervised fine-tuning, the performance now, since it learns from the real world, increases, let's say, to 70%, because now we have a realistic idea what's going on. And after we have generated all of this, we can reflect on it. Let's say we have an agent trajectory where we said, okay, we selected here some sold-out flight. Guess what? We will get an error state. And we say, hey, retry the selection. The page will come up and say, hey, please start all over again. So if we have now this feedback from the environment, we do have a reflection, a self-reflection generated by the LLM and says, okay, this was suboptimal. The selection ignored the availability at first. So what is the lesson the AI learned? Query the flight status API before choosing to avoid the errors. Or an alternative, activate, for example, a filter. Only flights that are available for my region. Having this reflection, do you know what we get? Do you see? We're not working anymore with a reward function. We're not working with a binary reward function. We're not working with a reward scoring. We get here a direct natural language lesson. This is something beautiful we're going to build upon. And if you have all this data, we fine-tune here on the augmented data, on the original trajectory, and all the reflection that we just generated here, distilling now the lessons. Now, this here, in total, refines our decision-making. That's why we are adding now checks for edge cases, boosting out-of-domain success, because... Look at this. This is here not just a lesson, but this is here a general lesson that is not only valid for flights from San Francisco to New York or from San Francisco to Israel or San Francisco to Sydney. So, out-of-the-main success goes up 65%. Yeah, self-reflection, I think it's clear. Perfect instant of a contrastive principle distilling supervision that we have that the self-reflection provides. The agent is not just told its action was wrong and we have a binary reward signal, but we have now a generalizable rule. The query had the flight status API before choosing here this flight. So you see precisely why here self-reflection is shown to improve our out-of-domain generalization. But you see what we get? This system now learns with pure trial and error interacting within an environment. It gets generalizable rules. And guess what? This will be the input to the Stanford study now. Ace. Agenic context engineering. What is it? Now we do have a problem over there. Because whenever we have this in the past, we have a problem. And we call this problem a context collapse. And this is a phenomenon within any LLM when tasked with iteratively and monolithically rewriting here a long and complex context. a system prompt or the complete agent memory this ei has a problem because it does a catastrophically over summarization at one point in time it is not copying the complete long and complex context but suddenly the eye decide now it's time i do a massive summarization Now this causes a sharp and irreversible loss of all detailed, specific and accumulated knowledge that was in the system prompt or that was here in the memory of our agent. This is a problem. Here we have it. Context collapse, you see? Adaptation to tokens here in the context. Yeah, grow. And then the LLM just decide, hey, I make it shorter. I compress it. It builds up a little bit. Yeah, look, 18,000 tokens. And then we fall down to 122 tokens. So we lose almost 99% here of the information. But if you have this real-time in stream, you see the accuracy goes from 66 to 57. So maybe we as a user, if it's a black box AI, we don't even notice that this is happening, but we lose information. This is not acceptable. So what do we do? Yeah, if you want to see this in spaces here, so we have an accumulation phase, a trigger phase, a collapse phase, and then AI simply has amnesia. The agent now operating with this collapse content has effectively forgotten its previous learning and its performance plummets. This is exactly here where we say goodbye, ciao knowledge, I just forgot you. If we do now this new Stanford paper, we have to overcome this problem. And ACE was now architected specifically to solve this problem by eliminating here all LLM monolithic rewrites in the system inherently and externally. What we do instead? We have an incremental delta update phase. Simple. We say we have now multiple objects, like a reflector, I'm going to show you this in a minute, but just as a summary here, analyzes the new experience and distills here some insight. And then the curator receives this insight. And it does not give the command to the LLM, hey, rewrite now the complete playbook, and at the end add this concise insight from the reflector. So we just have a concatenation, because this fails in some time, we just have now a programmatic operation, a deterministic non-LLM action, action, add something in the section, XYZ, and now the content is the concise insert generated by the reflector. And non-LLM action is the solution to this. So here we are. Integrating now ACE, so we have an evolving context as a playbook. So ACE treats you the system prompt or the complete memory as an evolving playbook accumulating now with each and every context with the environment. new strategy accumulating our strategy via those modular roles and i've showed you the generator the reflect and the curator and we have now a system that has two major benefits it avoids the brevity bias and the context collapse we just talked about the context collapse but what the hell is this bias this brevity bias is a tendency for an automated prompt refinement system deep in our ai to converge on shorter more generic instruction while systematically discharging valuable domain specific details and you might say unbelievable again this is not possible again yes this was in the pre-training data of your llm give you an example before the bias happens we have this prompt after the bias we have this prompt You see all the details, formula, total liability, showholders, equity, the balance sheet is missing data, use a linear regression to extrapolate, but I had a disclaimer. Gone. You're a financial analyst. Have a nice day. So our optimizer, our LLM, has incorrectly decided that the second prompt is better because it's shorter, works on most simple examples, and maybe we only go from 66% performance to 57% performance, but it's much easier for the eye to perform. Great. We don't want this. Therefore, ACE is taking care about this. So what do we have? Complete ACE. This is the flowchart of ACE. We have an initial context. This is the basic prompt. Book a flight, searching, selecting and paying. Great. And then a generator role. The generator produces now a trajectory using the current context. Let's say attempts booking here San Francisco to London history, but fails because we have an invalid credit card like mine. Then the reflector comes in and says, hey, wait a minute. So I have to analyze now the failure. What happened? Error. The card format is invalid. I produce now an insight. Do you see the parallality to the first paper? And the insight is validate the input pattern. It has to be, I don't know, 16 digit and no spaces and only slashes or dashes or whatever. Then the curator comes in and says, okay, I understand that there's an error. I understand why the error happened. And now... I apply a delta update here to the memory, to the system prompt, to my internal, and I append now the playbook without a full rewrite. Because with a full rewrite, we're going to have a collapse. So therefore, the strategy is now, if you do the payment, card payment, you know, here I give you a piece of code, regex check your card. It has to be in this specific format before you submit your card information to the internet. And the more the system learns in the interaction with the environment, the larger your playbook grows incrementally. But it will preserve now all the details that it learned. Not in an LLM summarization, but I just imagine this in a list. You just append the list. With handle sold out. Simply re-query with plus one day. You see, those incremental delta updates, as simple as they sound, They solve a major problem we had with AI. And you see, this is, if you want a key innovation in this paper, also to prevent context collapse. What is simple methodology? But AI itself is not able to do it within the AI. So, instead of asking here in LLM to remember to validate the credit cards, like you might think this is the right instruction, during a full prompt rewrite, where all the details might get lost at some point, the curator adds here precise, structured, and a permanent piece of knowledge, a code, a regex, jcard, to the playbook. Great, that's it. And now let's add those two papers together. Let's build context for those two papers. So the first paper was about the early experience, the supervised fine tuning, and the second paper from Stanford was about ACE. And now I can build an inner loop, and with age, i built the outer loop so we have an agent in contact with an environment with a booking system and i don't have to explain anything i just provide 10 simple expert traces trajectories for the ai to start learning the basis and then i say go have fun with this interface learn the agent performs tactical on the ground learning by executing all its actions, observing the immediate outcomes. You see, this is kind of our reward, but we don't need a reward function for this because we do have some real reward. This generates here the raw data of what works and what does not work because we see it here live. Can we book the flight or not? And then the outer loop ACE reflects on this raw data that we just generated to distill now strategic higher level principles. So we have here, if you want, The beginning with a little bit of a supervised fine-tuning, then we have a self-learning process to get here the raw data of the experiment of the interaction with the environment. And now here in the second outer loop with ACE, we distill it now into a strategic high-level principles, into a generalization in natural language, where we can apply this here or other flights around the globe. The principles are then codified into the playbook by the curator. So easy. And we have both papers, not just the Stanford paper. So the solution from the early experience paper for meta superintelligence is how to generate dense, reward-free supervision from self-exploration, you know, our world models, and distill transferable lessons from the mistakes by a self-reflection methodology. Now, of course, this depends on the intelligence of your system. If you have a GPT-5 High or a GPT-5 Pro or... I don't know what you... whatever system you have. Invest in the intelligence because this is essential. Because this is now the input to ACE. Because this distilled transferable lessons from the interaction with the environment become now the base. Accumulate the lessons over time in a structured evolving playbook, see the memory, that avoids the catastrophic forgetting and depravity bias. So simple. So you see? we combined the insights from meta just two days ago with the insight from stanford five days ago isn't this nice so whenever you see here on x or anywhere some statement here on a single paper put the paper in the perspective in the environment it belongs have a thematic clustering of all the other papers i don't know a week before two weeks before two weeks after so you understand what is going on in the world Interestingly, this is a pure US paper, no? Stanford, San Manuva, Berkeley, Meta, Superintelligence, FAIR, and Ohio State. But you see, they're all working here more or less on the same topics. So what we have now, if you look at this in a synergetic way, how EE provides you to raw experimental data that ACE then structures into a lasting wisdom, quotation mark, creating here a virtuous cycle of self-improvement for our agents. Now this makes sense, because we have access to the environment, and this is here our flight booking website, for example. So what I've shown you here is simple, the latest research in agent self-evolution. The agent improves via an internal reflection, the critiques see its own trajectories, grounded in the external states, the outcome here from the environment, from the booking website, bridging here super-wise fine-tuning static data, and the trial and error we have from reinforcement learning. You can see this from a complexity viewpoint because now the reflection abounded here by order n per step, if n is your trajectory length, leveraging LLM pre-trained browse to infer patterns without an exhaustive search. So we have a reduced solution space, but more about this in a later video. And both papers tell us, hey, if we do this isolated A, we get about 10% improvement, and if we do isolated B, we get about 10% improvement. in our multi-turn task and it is scalable to real domains from flight booking to finance operation to whatever you have and i have now to combine both papers both codes and see if what is the overall total if i have here the synergy of both papers but let's look at the result of each single paper let's look at the performance data The agents collected interaction data from their own action, transforming future state into supervision via implicit world modeling. We have here the supervised fine-tuning on the state action state triplets. To ground here the dynamics prediction and the self-reflection in EE, the LLM generated critique of suboptimal projectors versus the expert traces, distilling therefore lessons out here for our policy refinement. And... This addresses here beautifully Supervised Fine Tuning limited diversity and or else inefficiency in long horizon unverifiable settings. Because what you do if you do not have here a reward function? And the authors did this here in the first paper here across eight environments. Beautiful, beautiful. Yields 10% success rate. About 10% out of domain generalization over Supervised Fine Tuning. and it scales with 50% less expert data and even enhances here the follow-up reinforcement learning warm-up starts with plus 6.5% here in the final success cases. So nice. ACE, second paper, generator, reflector, curator. We talked about it. Great. Preserves the domain-specific heuristic without any weight updates of our transformer layers. No tensor updates. And I'm playing here offline. This means a prompt optimization. and online this means memory adaptation mode and the gains are about 10 on agents matching gpt 4.1 with deep seek version 3.1 plus 8.6 on domain task finance and look at this 86 lower latency than baseline like gpa using unlabeled execution signals now this is nice look at this If we have a React methodology with GEPA and we have now the latest React methodology with ACE, the latency, look at this, minus 82%. Wow. Offline, online in both cases. Here, online even minus 91%. Rollouts significantly reduced, minus 75%. So this means also the costs go down. So nice. So nice. But careful, you have to have the data for ACE. Careful, this is not just that you activate it and it gives you immediately dose. Yeah, GEPA. Remember, I have this specific video here, the end of reinforcement learning, GEPA, new genetic AI. This was by MIT UC Berkeley, a genetic Pareto prompt optimizer that thoroughly incorporates natural language reflection to learn high-level rules from trial and error. As you see, it's always trial and error. The system has to learn. Self-evolving AI system. But now have a look at this. Here we have everything. In-context learning, then from DSPi, MyPro version 2, then Geeper, just showed you the video, and then our new methodology ACE in two cases. First, with ground truth labels and ACE without ground truth labels. Here we have a DeepSync version 3.1 as a base LLM, offline, beautiful. Just go with anything that you like or let's go with the average, okay? You see, compared here to the... BaseMall, DeepSea version 3.1, has 69.1%. InContext Learning provides a plus 0.5 percentage point jump. MyProDSPy plus 1.8 percentage jump. Geeper plus 3%. Ace with Ground Truth plus 12%. And without Ground Truth, and this is the real impressive one, plus 8% because it has contact to the environment. This is a jump in the performance that is really interesting. Now, this is a very specific benchmark for the App World Agent benchmark. So this was selected by the authors for a good reason, I suppose. So you understand what I want to tell you. Check it for your specific use case in your domain for your complexity of your tasks. If you want to see maybe another example, let's go a step further and let's increase the complexity. It's the same topic. I just want to give you a little bit more challenging task. So we have an agent initialized here via SuperOS fine tuning, a limited set of expert demonstration, 20 trajectories for a task that is now a little bit more complex. Book a round-trip flight from San Francisco JFK for two adults in this time frame under $500. And maybe you want to sit at the window or wherever, so you understand what we are doing. We're just increasing the complexity. So we have to, as a startup, example expert human trajectory here, state, Use a query with these preferences. AI takes an action. Another action. Filtered by price. Another action. Enter passenger information. Name. Payment. Card. Whatever. Booking page. Action. Confirm. Submit. Great. So you see, this is here the sample trajectory. And now we do exactly the same. We do have our implicit role modeling. The agent proposes now actions and fine-tunes on the observed outcomes to predict your states. So you have a rollout of a multi-lag international task. The proposed action might be query API from San Francisco to London Heathrow and then from London Heathrow to Sydney, dates, whatever, future state or partial results because there's a problem, there's an unavailability layover constraint. So another action, assume alternative routing via Frankfurt. Future state updated the option, but price exceed the budget. So fine tune now on all this what the AI learned here. in the interaction with the real environment. Fine-tune on the triplets that we have, state at point T, the action at point T, and the follow-up state at T plus one. Grounding in the dynamics, our API return null for invalid layover, our prices may vary, whatever, but we have triplets, we have now training data from the real interaction with the environment that we can fine-tune our agent now on this advanced data, enhanced data. Okay, and enables participation of uncertainty, improve the success to 70% by modeling transition like, hey, if one leg of this international transfer is unavailable, query alternatives with a plus one day buffer. Think about the hotel and you got it. And then you remember EE has the second part, self-reflection. So for any suboptimal trajectory, the LLM critiques now it by comparing it to the human expert. Agent trajectory is select multi-leg without a visa check. And you get an error because the booking failed, because the visa was required, let's say, for the entry to Sydney. This reflection takes place, says, okay, this was suboptimal, the AI, because it omitted here the visa eligibility. So we have a lesson learned by the AI for international query visa API with your specific country codes before doing here flight selection. Alternative, fill to the routes by user passport assumption if unspecified. So you have a path A forward or a path B forward. And the more paths you have, the more options you have, the more feedback you get here from the booking system. Yes, you do have a trial and error system, but it learns real fast. You have now augmented data and you can fine tune now your agent on those. Also distilling lessons for the reasoning path. Because more and more lessons you have, the better the system understands here the cause effect of reality or of your booking agent. Plus it boosts out of domain generalization. Let's say you want to add a hotel, integrate booking with accommodation API, and you reach here a better performance. Early experience trajectories inform now our ACE, our second step. A reflected lesson like, hey, don't forget to check the visas, becomes now a delta instruction in the playbook of ACE. guiding here subsequent rollouts for precise prediction. You see, those papers, these two papers work so beautifully together. Sometimes it's really a pity that the authors don't know that Meta did publish the paper and Stanford published the paper just days apart if they would have sit together and said, hey, come on, but hey, otherwise you would not have watched this beautiful video. So you see, if you see this data. it's so beautiful if you see the opportunities you have the chances you have to build on those insights okay early experience does kill the diversity exposing your multi-domain edges yes the next video will be about a graph representation for ace great there we have it ace evolves not a agent memory as a playbook via model roles we have to fight the collapse therefore we accumulate our strategies by multiple objects and agents incrementally And therefore, we can avoid the collapse. If you want to see this here in another example, see it with more details, with more strategy details, this is it. Otherwise, again, we have our beautiful loop. Early experience supplies scalable supervision, the future states from the complex rollouts, while ACE organizes this into a persistent structure in a generalization, the deltas for multidomain handling. Fostering thereby self-improvement of the complete AI system, for long horizon uncertain tasks without external labels and no reward functions. Isn't this beautiful? So you see, and now understanding this we can say ACE offers a flexible and efficient alternative to conventional model fine tuning as adapting now context is generally cheaper than updating your model weights. Especially if you have a black box and you are not allowed Or you have to pay extreme amount of money to update your GPT-5 here or to fine tune your GPT-5 whatsoever to your data. I hope you enjoyed it.

The synergy between Early Experience and Agentic Context Engineering (ACE) creates a powerful, two-loop architecture for autonomous AI self-improvement. The Early Experience paradigm acts as the agent's tactical sensory system, generating a continuous stream of raw, grounded learning signals by exploring alternative actions and observing their immediate, reward-free consequences. 

This raw experiential data then feeds into the Agentic Context Engineering framework, which functions as the agent's strategic cognitive brain. Here, a Reflector analyzes these experiences to distill high-level principles, and a Curator surgically integrates them into an evolving, detailed "playbook" - a permanent knowledge base of strategies and pitfalls that avoids the destructive traps of context collapse and brevity bias. 

This creates a virtuous cycle: the structured wisdom of the ACE playbook makes the agent's future explorations more intelligent and efficient, which in turn provides richer, more nuanced data to further refine the playbook, allowing an agent to evolve from simply reacting to its environment to strategically accumulating expertise.

@stanford @UCBerkeley 

All rights w/ authors:
Agentic Context Engineering: Evolving Contexts for Self-Improving
Language Models
Qizheng Zhang 1, Changran Hu 2, Shubhangi Upasani 2, Boyuan Ma 2, Fenglu Hong 2,
Vamsidhar Kamanuru 2, Jay Rainton 2, Chen Wu 2, Mengmeng Ji 2, Hanchen Li 3,
Urmish Thakker 2, James Zou 1, Kunle Olukotun 1,
from 
1 Stanford University 
2 SambaNova Systems, Inc. 
3 UC Berkeley

Pre-print #2:
"Agent Learning via Early Experience"
Kai Zhang1,3,◦,⋆, Xiangchao Chen3,⋆, Bo Liu2,⋆, Tianci Xue3,⋆, Zeyi Liao3,⋆, Zhihan Liu1,⋆, Xiyao Wang1,⋆,
Yuting Ning3,⋆, Zhaorun Chen1,⋆, Xiaohan Fu1, Jian Xie3, Yuxuan Sun3, Boyu Gou3, Qi Qi1, Zihang Meng1,
Jianwei Yang1, Ning Zhang1, Xian Li2, Ashish Shah1, Dat Huynh1, Hengduo Li1, Zi Yang1, Sara Cao1,
Lawrence Jang1, Shuyan Zhou1,⋆, Jiacheng Zhu1,⋆, Huan Sun3,⋆, Jason Weston2,⋆, Yu Su3,†, Yifan Wu1,†
from
1 Meta Superintelligence Labs, 
2 FAIR at Meta, 
3 The Ohio State University

#aiexplained  
#machinelearning  
#scienceexplained #stanforduniversity