4 ms·

I believe this kind of graph exploration is what we need to progress reasoning in AI. Plain LLMS will fail. The link has tons of good references, including the

I believe this kind of graph exploration is what we need to progress reasoning in AI. Plain LLMS will fail. The link has tons of good references, including the Zobrist hashing https://en.wikipedia.org/wiki/Zobrist_hashing https://en.wikipedia.org/wiki/Zobrist_hashing for game tables. We need to find a good hashing for language based state description so that graph exploration doesn't explode computationally. Another good read for Tree Search is Thinking Fast and Slow: https://arxiv.org/abs/1705.08439 https://arxiv.org/abs/1705.08439 and Teaching Large Language Models to Reason with Reinforcement Learning: https://arxiv.org/abs/2403.04642 https://arxiv.org/abs/2403.04642 comparing the MCTS approach to other current RL strategies.

anonymous-panda 3y ago
Would it be impossible to marry the two somehow? It’s hard for me to believe that the brain only uses a single technique for everything and likely has many different tools in its toolbox with a selector on top to know how to leverage each appropriately.
vjerancrnjak 3y ago
This looks too low level. What might be a step forward is a joint learning of the state representation with the search algorithm. Search algorithm explores the NN representation of the state for which you can get the cost. https://sites.google.com/view/genie-2024/ https://sites.google.com/view/genie-2024/ Genie from DeepMind is a good demonstration where discrete state is being modeled. NN learns a very complex representation with collision detection and actions. Instead of decoding that state into pixels, search could probably be done directly on that state. Of course, this architecture could be very different.
- emmett 3y ago
  Active Inference is kinda "a joint learning of the state representation with the search algorithm"...or at least, related to the idea. I like framing it as a joint learning problem.
andrewdb 3y ago
An over-simplified approach that would be interesting to explore: 1. Given a collection of logical arguments, figure out how to assign a hash to each argument. 2. Build a Merkle tree representing the argument's hashes, nesting according to first principles. 3. If an argument is challenged successfully, the argument's hash changes, rendering descendent argument hashes invalid.

anonymous-panda 3y ago

Would it be impossible to marry the two somehow? It’s hard for me to believe that the brain only uses a single technique for everything and likely has many different tools in its toolbox with a selector on top to know how to leverage each appropriately.

vjerancrnjak 3y ago

This looks too low level. What might be a step forward is a joint learning of the state representation with the search algorithm. Search algorithm explores the NN representation of the state for which you can get the cost. https://sites.google.com/view/genie-2024/ https://sites.google.com/view/genie-2024/ Genie from DeepMind is a good demonstration where discrete state is being modeled. NN learns a very complex representation with collision detection and actions. Instead of decoding that state into pixels, search could probably be done directly on that state. Of course, this architecture could be very different.

emmett 3y ago
Active Inference is kinda "a joint learning of the state representation with the search algorithm"...or at least, related to the idea. I like framing it as a joint learning problem.

andrewdb 3y ago

An over-simplified approach that would be interesting to explore: 1. Given a collection of logical arguments, figure out how to assign a hash to each argument. 2. Build a Merkle tree representing the argument's hashes, nesting according to first principles. 3. If an argument is challenged successfully, the argument's hash changes, rendering descendent argument hashes invalid.