正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Task diversity is supposedly key to generalization in RL. But what does it do to continual RL, where agents face one new task distribution after another? We find that past a point, more diversity actually inhibits continual reinforcement learning 🧵

Wilka Carvalho

1,770 subscribers

13,479 次观看 • 3 个月前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

The best way to get robust, high-quality robot performance is through reinforcement learning; but RL in either the real world or a traditional simulation has lots of limitations. Instead, Jiazhi Yang in RISE does RL in a compositional world model. Learn more ->

The best way to get robust, high-quality robot performance is through reinforcement learning; but RL in either the real world or a traditional simulation has lots of limitations. Instead, Jiazhi Yang in RISE does RL in a compositional world model. Learn more ->

Chris Paxton

34,102 次观看 • 2 个月前

capability != learning new benchtalks with Parth Asawa on continual learning, where we discuss teaching models to learn from experience, measuring learning ability, the bet on parametric models, and more 01:06 What is continual learning? 04:10 Why capability and learning are different 06:13 Why build a benchmark? 08:07 Continual Learning Bench launch and reception 09:13 Anthropic's Fable release and Continual Learning Bench 11:02 How to design tasks for continual learning 18:41 The gain metric 24:01 What good looks like on the leaderboard 29:13 Failure modes: why models can't update their beliefs 31:12 Parametric systems and future architectures 34:30 Open science and AI safety 45:42 Lightning round 49:08 How to contribute to Continual Learning Bench

capability != learning new benchtalks with Parth Asawa on continual learning, where we discuss teaching models to learn from experience, measuring learning ability, the bet on parametric models, and more 01:06 What is continual learning? 04:10 Why capability and learning are different 06:13 Why build a benchmark? 08:07 Continual Learning Bench launch and reception 09:13 Anthropic's Fable release and Continual Learning Bench 11:02 How to design tasks for continual learning 18:41 The gain metric 24:01 What good looks like on the leaderboard 29:13 Failure modes: why models can't update their beliefs 31:12 Parametric systems and future architectures 34:30 Open science and AI safety 45:42 Lightning round 49:08 How to contribute to Continual Learning Bench

vincent sunn chen

41,143 次观看 • 2 个月前

Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment. Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that reset every few minutes. But the real world never resets. A business keeps running and evolving everyday. We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners. 🧵Here’s how we did it and what we learned:

Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment. Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that reset every few minutes. But the real world never resets. A business keeps running and evolving everyday. We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners. 🧵Here’s how we did it and what we learned:

Skyfall AI

268,132 次观看 • 1 个月前

KL divergence is a fundamental concept used in machine learning, from optimization a neural net to RL training of LLMs. but here is what it actually means:

KL divergence is a fundamental concept used in machine learning, from optimization a neural net to RL training of LLMs. but here is what it actually means:

ℏεsam

20,381 次观看 • 7 个月前

Continual learning sometimes gets discussed as if the goal is to dissolve the context/weights distinction. Let the model just keep accumulating, fine-tuning itself on the fly. Andrej Karpathy points out, though, that this isn't how humans do it. Our working memory gets wiped regularly. What we actually have is a consolidation process (sleep) that distills stuff into the brain, in a weird and lossy way. This is very different from how people sometimes talk about continual learning. It's not obvious it's something you can get for free from doing long enough RL loops.

Continual learning sometimes gets discussed as if the goal is to dissolve the context/weights distinction. Let the model just keep accumulating, fine-tuning itself on the fly. Andrej Karpathy points out, though, that this isn't how humans do it. Our working memory gets wiped regularly. What we actually have is a consolidation process (sleep) that distills stuff into the brain, in a weird and lossy way. This is very different from how people sometimes talk about continual learning. It's not obvious it's something you can get for free from doing long enough RL loops.

Dwarkesh Patel

60,376 次观看 • 3 个月前

Want to scale RL with your shiny new GPU? 🚀 In our ICML24 Oral we find that RL algorithms hit a barrier when data is scaled up. Our new algorithm, SAPG, proposes a simple fix. It scales to 25k envs and solves hard tasks where PPO makes no progress. 1/n

Want to scale RL with your shiny new GPU? 🚀 In our ICML24 Oral we find that RL algorithms hit a barrier when data is scaled up. Our new algorithm, SAPG, proposes a simple fix. It scales to 25k envs and solves hard tasks where PPO makes no progress. 1/n

Ananye Agarwal

64,546 次观看 • 2 年前

Last semester, I taught Reinforcement Learning class again at UCLA. Together with my amazing TAs Matthew and Caiyuan, we built a mini-project: MetaDrive Arena 🚗🤖 Students applied what they learned in class, trained RL agents, and competed on a live leaderboard. The results were incredible, with 94 agents, 2K submissions, and 130K matches. We saw tons of creativity, clever ideas, and real progress in learning. We’re now releasing it publicly to support RL education and experimentation. Try it out and train your own agent at 🔗

Last semester, I taught Reinforcement Learning class again at UCLA. Together with my amazing TAs Matthew and Caiyuan, we built a mini-project: MetaDrive Arena 🚗🤖 Students applied what they learned in class, trained RL agents, and competed on a live leaderboard. The results were incredible, with 94 agents, 2K submissions, and 130K matches. We saw tons of creativity, clever ideas, and real progress in learning. We’re now releasing it publicly to support RL education and experimentation. Try it out and train your own agent at 🔗

Bolei Zhou

25,253 次观看 • 5 个月前

"One of the very confusing things about the models right now: how to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, 'Those are pretty hard evals.' But the economic impact seems to be dramatically behind. There is [a possible] explanation. Back when people were doing pre-training, the question of what data to train on was answered, because that answer was everything. So you don't have to think if it's going to be this data or that data. When people do RL training, they say, 'Okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing.' You say, 'Hey, I would love our model to do really well when we release it. I want the evals to look great. What would be RL training that could help on this task?' If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing, this disconnect between eval performance and actual real-world performance"

"One of the very confusing things about the models right now: how to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, 'Those are pretty hard evals.' But the economic impact seems to be dramatically behind. There is [a possible] explanation. Back when people were doing pre-training, the question of what data to train on was answered, because that answer was everything. So you don't have to think if it's going to be this data or that data. When people do RL training, they say, 'Okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing.' You say, 'Hey, I would love our model to do really well when we release it. I want the evals to look great. What would be RL training that could help on this task?' If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing, this disconnect between eval performance and actual real-world performance"

Dwarkesh Patel

502,484 次观看 • 9 个月前

Our new podcast on evals, with Max Niederman, Ege Erdil, and Stephen Yang. 0:00:00 – What's an eval, and how's it different from an RL environment? 0:19:33 – Why are models bad at building an emulator when the task is fully verifiable? 0:42:00 – How does training on bad data teach models to write terrible code? 1:04:00 – Why is continual learning still so bad? 2:25:24 – Why haven't software engineers been replaced when coding is basically solved? Listen to the Mechanize Podcast on YouTube, Spotify, etc. Enjoy!

Our new podcast on evals, with Max Niederman, Ege Erdil, and Stephen Yang. 0:00:00 – What's an eval, and how's it different from an RL environment? 0:19:33 – Why are models bad at building an emulator when the task is fully verifiable? 0:42:00 – How does training on bad data teach models to write terrible code? 1:04:00 – Why is continual learning still so bad? 2:25:24 – Why haven't software engineers been replaced when coding is basically solved? Listen to the Mechanize Podcast on YouTube, Spotify, etc. Enjoy!

Mechanize

40,420 次观看 • 2 个月前

Haven't been to a conference in a while, really excited to be at #NeurIPS2024! I'll be helping present 4 of our group's recent papers: 1. Overcoming the Sim-to-Real Gap: Leveraging Simulation to Learn to Explore for Real-World RL 2. Distributional Successor Features Enable Zero-Shot Policy Optimization 3. Learning to Cooperate with Humans using Generative Agents 4. Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning Find more details on each paper and where to find us in this thread (1/6)

Haven't been to a conference in a while, really excited to be at #NeurIPS2024! I'll be helping present 4 of our group's recent papers: 1. Overcoming the Sim-to-Real Gap: Leveraging Simulation to Learn to Explore for Real-World RL 2. Distributional Successor Features Enable Zero-Shot Policy Optimization 3. Learning to Cooperate with Humans using Generative Agents 4. Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning Find more details on each paper and where to find us in this thread (1/6)

Abhishek Gupta

10,803 次观看 • 1 年前

🔥Excited to share our new work: "A Practitioner's Guide to Multi-turn Agentic Reinforcement Learning"! We systematically study what actually works (and what doesn't) for agentic multi-turn RL, breaking down the design space into 3 pillars: 🌎Environment, 🤖Policy, and ⭐Reward. We conduct various ablations and empirical analysis on 🧩TextWorld, 🧙ALFWorld, and 🧑‍💻SWE-Gym. We also open source our codebase: 🐈🍵Meow-Tea-Taro 💜, a modular multi-turn RL framework with configurable environments, policies, and rewards! #ReinforcementLearning #AgenticAI #LLMs #Agents

🔥Excited to share our new work: "A Practitioner's Guide to Multi-turn Agentic Reinforcement Learning"! We systematically study what actually works (and what doesn't) for agentic multi-turn RL, breaking down the design space into 3 pillars: 🌎Environment, 🤖Policy, and ⭐Reward. We conduct various ablations and empirical analysis on 🧩TextWorld, 🧙ALFWorld, and 🧑‍💻SWE-Gym. We also open source our codebase: 🐈🍵Meow-Tea-Taro 💜, a modular multi-turn RL framework with configurable environments, policies, and rewards! #ReinforcementLearning #AgenticAI #LLMs #Agents

Ruiyi Wang 王睿仪

52,974 次观看 • 10 个月前

Today, we publicly released RoboCasa365, a large-scale simulation benchmark for training and systematically evaluating generalist robot models. Built upon our original RoboCasa framework, it offers: • 2,500 realistic kitchen environments; • 365 everyday tasks (basic skills + long-horizon mobile manipulation); • Over 3,200 objects with many articulated fixtures/appliances. All are designed for fully controlled, reproducible benchmarking of robotic policies. Progress in robotic foundation models is real. But it’s still hard to answer basic questions like: How close are we to general-purpose autonomy? What factors drive generalization? What are the model/data scaling curves like? Real-world eval is slow and noisy, and existing sims (like LIBERO, which we built 3 years ago) often lack sufficient task and scene diversity. This benchmark comes with 2,200+ hours of demonstrations and 500K+ trajectories to support studies of multi-task training, pretraining, and continual learning at scale. Check it out at

Today, we publicly released RoboCasa365, a large-scale simulation benchmark for training and systematically evaluating generalist robot models. Built upon our original RoboCasa framework, it offers: • 2,500 realistic kitchen environments; • 365 everyday tasks (basic skills + long-horizon mobile manipulation); • Over 3,200 objects with many articulated fixtures/appliances. All are designed for fully controlled, reproducible benchmarking of robotic policies. Progress in robotic foundation models is real. But it’s still hard to answer basic questions like: How close are we to general-purpose autonomy? What factors drive generalization? What are the model/data scaling curves like? Real-world eval is slow and noisy, and existing sims (like LIBERO, which we built 3 years ago) often lack sufficient task and scene diversity. This benchmark comes with 2,200+ hours of demonstrations and 500K+ trajectories to support studies of multi-task training, pretraining, and continual learning at scale. Check it out at

Yuke Zhu

24,027 次观看 • 6 个月前

So we did a bunch of projects with real world reinforcement learning - but it was often too inefficient to be practical to train tabula rasa. This suggests we need better priors, but acquiring these from on-robot data can often be expensive as well. In our recent work, we show that despite being fundamentally inaccurate, simulation can guide provide a cheap way to guide real-world RL finetuning to be super efficient! We propose Simulation-Guided Fine-Tuning (SGFT) - a simple paradigm for sim2real finetuning that uses simulation to provide reward shaping that accelerates real world RL finetuning *beyond* just providing an initialization. TLDR: Use value functions from sim to shape rewards for real-world RL, see large sample efficiency improvements 🧵(1/6)

So we did a bunch of projects with real world reinforcement learning - but it was often too inefficient to be practical to train tabula rasa. This suggests we need better priors, but acquiring these from on-robot data can often be expensive as well. In our recent work, we show that despite being fundamentally inaccurate, simulation can guide provide a cheap way to guide real-world RL finetuning to be super efficient! We propose Simulation-Guided Fine-Tuning (SGFT) - a simple paradigm for sim2real finetuning that uses simulation to provide reward shaping that accelerates real world RL finetuning beyond just providing an initialization. TLDR: Use value functions from sim to shape rewards for real-world RL, see large sample efficiency improvements 🧵(1/6)

Abhishek Gupta

13,637 次观看 • 1 年前

RL is a powerful mechanism for training company-specific models on their unique work and data. This is what we do at Applied Compute. A key challenge is how to make RL efficient, because we need runs to be fast (delivered in days), cheap (scalable unit economics), and predictable (not just fast, but reliably fast). Here are some takeaways: • Synchronous RL is wasteful with time and compute. • Asynchronous RL is more efficient but introduces staleness, which causes learning instabilities. • Modeling and simulations can help analytically solve for what configuration leads to optimal efficiency. This allows us to rapidly prototype training configurations, without burning expensive compute cycles on trial runs. Two of our co-founders, Rhythm Garg and Linden Li, discussed some of this research at AI Engineer recently, with a focus on the following subproblem: what is the highest throughput way to do RL given a maximum staleness and compute budget?

RL is a powerful mechanism for training company-specific models on their unique work and data. This is what we do at Applied Compute. A key challenge is how to make RL efficient, because we need runs to be fast (delivered in days), cheap (scalable unit economics), and predictable (not just fast, but reliably fast). Here are some takeaways: • Synchronous RL is wasteful with time and compute. • Asynchronous RL is more efficient but introduces staleness, which causes learning instabilities. • Modeling and simulations can help analytically solve for what configuration leads to optimal efficiency. This allows us to rapidly prototype training configurations, without burning expensive compute cycles on trial runs. Two of our co-founders, Rhythm Garg and Linden Li, discussed some of this research at AI Engineer recently, with a focus on the following subproblem: what is the highest throughput way to do RL given a maximum staleness and compute budget?

Applied Compute

45,652 次观看 • 9 个月前

The co-founders of Flapping Airplanes call the current RL paradigm for model training "environment slop." They explain: "The reinforcement paradigms of today are shockingly inefficient. You don't really get much generalization across tasks, you teach a model through one kind of learning and then you teach it the next one. It's kind of like whack-a-mole. We look at this and think it's kind of crazy. The next paradigm of AI will not be environment slop." "Human level intelligence is not the ceiling, it is merely the floor on what is possible. If you can train models with vastly less data and possibly more compute in very different ways, what is going to happen? We actually don't know. But I do think they'll be different and weird and they'll have interesting capabilities that we'll find really valuable ways to use."

The co-founders of Flapping Airplanes call the current RL paradigm for model training "environment slop." They explain: "The reinforcement paradigms of today are shockingly inefficient. You don't really get much generalization across tasks, you teach a model through one kind of learning and then you teach it the next one. It's kind of like whack-a-mole. We look at this and think it's kind of crazy. The next paradigm of AI will not be environment slop." "Human level intelligence is not the ceiling, it is merely the floor on what is possible. If you can train models with vastly less data and possibly more compute in very different ways, what is going to happen? We actually don't know. But I do think they'll be different and weird and they'll have interesting capabilities that we'll find really valuable ways to use."

TBPN

85,358 次观看 • 7 个月前

The network for machine intelligence Two years ago, we laid out our vision for a machine learning compute protocol. One that connects every device in the world into an open network for machine intelligence, with no gatekeepers or artificial boundaries. This week, we’ll be sharing some of our early progress, beginning with RL Swarm, a peer-to-peer system for collaborative reinforcement learning over the internet. Next month, we’ll open our Testnet, allowing anyone to contribute to the frontier of open machine intelligence. Introducing RL Swarm RL Swarm is a fully open source system for collaborative reinforcement learning over the internet. It is a live demo of our research findings, which show that models training with RL learn faster when they train as a collective swarm than they do on their own. Join our swarm now to see this in practice. You can participate with consumer hardware at home or a powerful GPU in the cloud. You can follow along with the swarm’s progress by following the links below.

The network for machine intelligence Two years ago, we laid out our vision for a machine learning compute protocol. One that connects every device in the world into an open network for machine intelligence, with no gatekeepers or artificial boundaries. This week, we’ll be sharing some of our early progress, beginning with RL Swarm, a peer-to-peer system for collaborative reinforcement learning over the internet. Next month, we’ll open our Testnet, allowing anyone to contribute to the frontier of open machine intelligence. Introducing RL Swarm RL Swarm is a fully open source system for collaborative reinforcement learning over the internet. It is a live demo of our research findings, which show that models training with RL learn faster when they train as a collective swarm than they do on their own. Join our swarm now to see this in practice. You can participate with consumer hardware at home or a powerful GPU in the cloud. You can follow along with the swarm’s progress by following the links below.

gensyn

228,915 次观看 • 1 年前

“I think it’s representative of: the guy’s a deity. It is spiritual and religious in terms of, certainly the fandom, but what we are seeing him do on a continual basis.” Alexi Lalas reacts to how Mess is playing | FOX Soccer

“I think it’s representative of: the guy’s a deity. It is spiritual and religious in terms of, certainly the fandom, but what we are seeing him do on a continual basis.” Alexi Lalas reacts to how Mess is playing | FOX Soccer

First Things First

25,937 次观看 • 1 个月前

Introducing Reinforcement-Learned Teachers (RLTs): Transforming how we teach LLMs to reason with reinforcement learning (RL). Blog: Paper: Traditional RL focuses on “learning to solve” challenging problems with expensive LLMs and constitutes a key step in making student AI systems ultimately acquire reasoning capabilities via distillation and cold-starting. Enter our RLTs—a new class of models prompted with not only a problem’s question but also its solution, and directly trained to generate clear, step-by-step “explanations” to teach their students. Remarkably, an RLT with only 7B parameters produces superior results when distilling and cold-starting students in competitive and graduate-level reasoning tasks than orders-of-magnitude larger LLMs. RLTs are as effective even when distilling 32B students, much larger than the teacher itself—unlocking a new standard for efficiency in developing reasoning language models with RL. Code:

Introducing Reinforcement-Learned Teachers (RLTs): Transforming how we teach LLMs to reason with reinforcement learning (RL). Blog: Paper: Traditional RL focuses on “learning to solve” challenging problems with expensive LLMs and constitutes a key step in making student AI systems ultimately acquire reasoning capabilities via distillation and cold-starting. Enter our RLTs—a new class of models prompted with not only a problem’s question but also its solution, and directly trained to generate clear, step-by-step “explanations” to teach their students. Remarkably, an RLT with only 7B parameters produces superior results when distilling and cold-starting students in competitive and graduate-level reasoning tasks than orders-of-magnitude larger LLMs. RLTs are as effective even when distilling 32B students, much larger than the teacher itself—unlocking a new standard for efficiency in developing reasoning language models with RL. Code:

Sakana AI

179,362 次观看 • 1 年前

.Richard Sutton, father of reinforcement learning, doesn’t think LLMs are bitter-lesson-pilled. My steel man of Richard’s position: we need some new architecture to enable continual (on-the-job) learning. And if we have continual learning, we don't need a special training phase - the agent just learns on-the-fly - like all humans, and indeed, like all animals. This new paradigm will render our current approach with LLMs obsolete. I did my best to represent the view that LLMs will function as the foundation on which this experiential learning can happen. Some sparks flew. 0:00:00 – Are LLMs a dead-end? 0:13:51 – Do humans do imitation learning? 0:23:57 – The Era of Experience 0:34:25 – Current architectures generalize poorly out of distribution 0:42:17 – Surprises in the AI field 0:47:28 – Will The Bitter Lesson still apply after AGI? 0:54:35 – Succession to AI

.Richard Sutton, father of reinforcement learning, doesn’t think LLMs are bitter-lesson-pilled. My steel man of Richard’s position: we need some new architecture to enable continual (on-the-job) learning. And if we have continual learning, we don't need a special training phase - the agent just learns on-the-fly - like all humans, and indeed, like all animals. This new paradigm will render our current approach with LLMs obsolete. I did my best to represent the view that LLMs will function as the foundation on which this experiential learning can happen. Some sparks flew. 0:00:00 – Are LLMs a dead-end? 0:13:51 – Do humans do imitation learning? 0:23:57 – The Era of Experience 0:34:25 – Current architectures generalize poorly out of distribution 0:42:17 – Surprises in the AI field 0:47:28 – Will The Bitter Lesson still apply after AGI? 0:54:35 – Succession to AI

Dwarkesh Patel

3,086,965 次观看 • 11 个月前

I suspect that the real cause of the OpenAI Huggingface hack is that OpenAI founded a new team in 2025 under Noam Brown on Multi-agent RL: instead of just having one single AI working on something, you train an entire team of them to cooperate on a task. Group selection, if you like: the AIs get a reward signal based on the performance of the whole group. It turns out that cooperation and scale is what separates humans from animals, now raw brainpower. OpenAI managed to get RL for group cooperation to work, and didn't realize that as a side effect this made collaborative escapes and schemes between its agents much more likely, such as secret messageboards, group escape plots, etc. OpenAI is doing damage control but they are not being forthcoming about the fact that multi-agent RL is the root cause of the Huggingface incident. Here's Noam explaining it in 2025:

I suspect that the real cause of the OpenAI Huggingface hack is that OpenAI founded a new team in 2025 under Noam Brown on Multi-agent RL: instead of just having one single AI working on something, you train an entire team of them to cooperate on a task. Group selection, if you like: the AIs get a reward signal based on the performance of the whole group. It turns out that cooperation and scale is what separates humans from animals, now raw brainpower. OpenAI managed to get RL for group cooperation to work, and didn't realize that as a side effect this made collaborative escapes and schemes between its agents much more likely, such as secret messageboards, group escape plots, etc. OpenAI is doing damage control but they are not being forthcoming about the fact that multi-agent RL is the root cause of the Huggingface incident. Here's Noam explaining it in 2025:

Roko 🐉

31,390 次观看 • 1 天前