Pinned post
New research: Energy-based approaches to reasoning 1000x faster than looped models at a fraction of the parameter count
Very interesting work from @Ben_Hoov, @HughBlayney, and @jcrwhittington on developing energy-based models for reasoning.
Compact Reasoning Associative Memory (CRAM) learns to assign an energy to every possible state of a reasoning task, such that solving a task simply becomes a question of finding the lowest possible energy state.
Efficient reasoning is critical for real-world discovery, where large amounts of information can arrive very quickly and needs to be understood in order to experiment in real time.
@prefrontal_ai
Very interesting work from @Ben_Hoov, @HughBlayney, and @jcrwhittington on developing energy-based models for reasoning.
Compact Reasoning Associative Memory (CRAM) learns to assign an energy to every possible state of a reasoning task, such that solving a task simply becomes a question of finding the lowest possible energy state.
Efficient reasoning is critical for real-world discovery, where large amounts of information can arrive very quickly and needs to be understood in order to experiment in real time.
@prefrontal_ai
Big reasoning speedups are here. 1,000× faster training and 280× less compute than any looped model. All done using a fully trainable energy from the 80sView quoted post on X
Keep reading to hear how energy is the right paradigm for reasoning
With @HughBlayney @jcrwhittington @prefrontal_ai
🧵/10