Looped transformers tie weights across layers, allowing the network to spend more iterations on harder inputs. Fixed-Point Reasoners (FPRM) achieve this adaptive computation with a damped fixed-point solver, but they operate in an unbounded Euclidean space. A new arXiv paper, "Spherical Fixed-Point Reasoners: Adaptive Loop Halting on the Hypersphere," proposes a variant that instead works on a hypersphere.
By moving the fixed-point iteration to a hypersphere, the method changes the underlying geometry of the solver. The provided abstract does not include experimental results, so the practical benefits are not yet detailed. The work is positioned as an alternative to the unbounded Euclidean setting of previous FPRM models, with the halting decision now tied to spherical coordinates.