Loading video...
Video Failed to Load
What’s left w/ foundation models? We found that they still can't ground modular concepts across domains. We present Logic-Enhanced FMs:🤝FMs & neuro-symbolic concept learners. We learn abstractions of concepts like “left” across domains & do domain-independent reasoning w/ LLMs.
49,981 views • 2 years ago •via X (Twitter)
7 Comments
Notable prior works have proposed LLMs for reasoning with execution from pretrained VLMs, but they are inference-only and can't be made trainable. Our model (LEFT) can learn new concept grounding from data in domains wo/ predefined models, as its executor is fully differentiable.
LEFT leverages LLMs to take language queries and output programs in a general first-order logic reasoning language, shared across domains and tasks. LEFT's executor then executes the programs with learnable domain-specific grounding modules, initialized with LLM-parsed concepts.
We can do the same general decomposition and execution in a variety of domains and for a variety of tasks (see more examples on our project page). Concepts in language serve as abstractions that enable such generalization.
The unified LEFT framework can perform visual reasoning in 2D, 3D, temporal motion, and robotic manipulation domains. It can also zero-shot transfer its concept knowledge to unseen tasks, through flexible LLM-generated logic & effective reuse of learned, modular visual concepts.
We propose Logic-Enhanced FMs as a general framework for concept learning & reasoning across domains and tasks. LEFT does not require predefined programs for new datasets and is easy to build on. We release demos to show how to apply LEFT on a new dataset in ~100 lines of code!
Excited to present this work at @NeurIPSConf with the wonderful @maojiayuan, Josh Tenenbaum, and @jiajunwu_cs! Paper: Website: Code:
Hey is there a chance I can dm you ?
![.Andrej Karpathy says that LLMs currently lack the cultural accumulation and self-play that propelled humans out of the savannah: Culture: > “Why can’t an LLM write a book for the other LLMs? Why can’t other LLMs read this LLM’s book and be inspired by it, or shocked by it?” Self play: > “It’s extremely powerful. Evolution has a lot of competition driving intelligence and evolution. AlphaGo is playing against itself and that’s how it learns to get really good at Go. There’s no equivalent of self-play in LLMs. Why can’t an LLM, for example, create a bunch of problems that another LLM is learning to solve? Then the LLM is always trying to serve more and more difficult problems.” I asked Karpathy why LLMs still aren't yet able to build up culture the way humans do. > “The dumber models remarkably resemble a kindergarten student. [The smartest models still feel like] elementary school students though. Somehow, we still haven’t graduated enough where [these models] can take over. My Claude Code or Codex, they still feel like this elementary-grade student. I know that they can take PhD quizzes, but they still cognitively feel like a kindergarten.” > “I don’t think they can create culture because they’re still kids. They’re savant kids. They have perfect memory. They can convincingly create all kinds of slop that looks really good. But I still think they don’t really know what they’re doing. They don’t really have the cognition across all these little checkboxes that we still have to collect.”](https://image.24vids.com/tw-1980333945385562176/amplify_video_thumb/1980333012060409856/img/1iuQ3tw9AOukO4E0.jpg)
![Wilfrid Mbappé about Kylian Mbappé and their family: M: “ Do you think we dehumanised Kylian? ” W: “ No, not at all. I think that he protected himself like i did. At some point you put a shell around yourself. and people only see your shell, because that’s the only thing we show, they don’t see whats inside. We keep our humanity inside, because if people break through or shell, what’s left for us? We are just ordinary people so we keep that humanity and when we come home, close the door, we do silly things, we laugh, we tease each other, we joke around like everyone else. We have to keep that for ourselves and that’s what we, as parents, did, to preserve. M: “ Who teases others the most in your family? ” W: “ All of us tease each other a lot but i would say it’s me 😂. Teasing has been our driving force too. When Kylian has a bad game, i get him, i tease him about his match, i joke around. I try to lighten up the mood because it’s just a sport. […] The football has become some kind of a social success for some people and i’m trying to fight against this idea. You can’t judge people based on that. You don’t know them, you don’t know their family and their problems, situation they are in etc. People shouldn’t judge others when they don’t know anything about their life, it’s easy to do from the outside not knowing anything and people use it as excuse. We can all do it. You need to get to know the person first, to judge someone. That’s why i don’t judge people. ”](https://image.24vids.com/tw-2021875788413305226/amplify_video_thumb/2021875470493532160/img/50cRtTqVTbljxU-J.jpg)
![The most interesting part for me is where Andrej Karpathy describes why LLMs aren't able to learn like humans. As you would expect, he comes up with a wonderfully evocative phrase to describe RL: “sucking supervision bits through a straw.” A single end reward gets broadcast across every token in a successful trajectory, upweighting even wrong or irrelevant turns that lead to the right answer. > “Humans don't use reinforcement learning, as I've said before. I think they do something different. Reinforcement learning is a lot worse than the average person thinks. Reinforcement learning is terrible. It just so happens that everything that we had before is much worse.” So what do humans do instead? > “The book I’m reading is a set of prompts for me to do synthetic data generation. It's by manipulating that information that you actually gain that knowledge. We have no equivalent of that with LLMs; they don't really do that.” > “I'd love to see during pretraining some kind of a stage where the model thinks through the material and tries to reconcile it with what it already knows. There's no equivalent of any of this. This is all research.” Why can’t we just add this training to LLMs today? > “There are very subtle, hard to understand reasons why it's not trivial. If I just give synthetic generation of the model thinking about a book, you look at it and you're like, 'This looks great. Why can't I train on it?' You could try, but the model will actually get much worse if you continue trying.” > “Say we have a chapter of a book and I ask an LLM to think about it. It will give you something that looks very reasonable. But if I ask it 10 times, you'll notice that all of them are the same.” > “You're not getting the richness and the diversity and the entropy from these models as you would get from humans. How do you get synthetic data generation to work despite the collapse and while maintaining the entropy? It is a research problem.” How do humans get around model collapse? > “These analogies are surprisingly good. Humans collapse during the course of their lives. Children haven't overfit yet. They will say stuff that will shock you. Because they're not yet collapsed. But we [adults] are collapsed. We end up revisiting the same thoughts, we end up saying more and more of the same stuff, the learning rates go down, the collapse continues to get worse, and then everything deteriorates.” In fact, there’s an interesting paper arguing that dreaming evolved to assist generalization, and resist overfitting to daily learning - look up The Overfitted Brain by Erik Hoel. I asked Karpathy: Isn’t it interesting that humans learn best at a part of their lives (childhood) whose actual details they completely forget, adults still learn really well but have terrible memory about the particulars of the things they read or watch, and LLMs can memorize arbitrary details about text that no human could but are currently pretty bad at generalization? > “[Fallible human memory] is a feature, not a bug, because it forces you to only learn the generalizable components. LLMs are distracted by all the memory that they have of the pre-trained documents. That's why when I talk about the cognitive core, I actually want to remove the memory. I'd love to have them have less memory so that they have to look things up and they only maintain the algorithms for thought, and the idea of an experiment, and all this cognitive glue for acting.”](https://image.24vids.com/tw-1979259041013731752/amplify_video_thumb/1979257571786543104/img/OqUh56yit9gUQWnb.jpg)
