Definition
Plain language
A model design that finds an answer by jumping straight to where its internal computation would settle, instead of iterating to get there.
As stated in the literature
An architecture that frames recurrent refinement as a fixed-point problem and uses implicit differentiation for constant-memory training, with the equilibrium living in output-embedding space.
Also called: Attractor Models
Why it matters: It promises the benefits of deep iterative reasoning at the cost of a single forward pass, with constant memory regardless of how deep the implicit reasoning is.
For example, instead of running ten refinement steps to settle an answer, the Attractor Model jumps directly to the fixed point those steps would converge to.
Heard on the show
“The paper is "Solve the Loop: Attractor Models for Language and Reasoning," by Jacob Fein-Ashley and Paria Rashidinejad at the University of Southern California.”Episode 041 — When the Iteration Teaches the Model to Skip the Iteration