POLYDIM
A novel reasoning architecture based on high-dimensional manifold representations. It maps reasoning trajectories to isometric geodesics on a sphere, avoiding the information loss caused by one-dimensional token serialization.
Timeline (last 90 days)
A research group on Reddit has publicly released their work POLYDIM / EinsofOS, which attempts to circumvent the quadratic complexity of attention and the information loss of one-dimensional tokens at the representation level.
The core claim of POLYDIM is: rather than using discrete one-dimensional autoregressive tokens to carry reasoning, it maps reasoning trajectories directly onto isometric geodesics on the sphere S^(D-1), where D ≥ 1024.
The paper optimizes on the Stiefel manifold St(k, D), using a fourth-order polar decomposition retraction together with a Neumaier compensated accumulator, and reports an orthogonality error of ||X^T X - I_k|| ≤ 10^-15 with no catastrophic drift.
This orthogonality error bound (10^-15) is the part the authors actively invite community criticism on, and the authors explicitly state that they hope to receive challenges regarding the boundaries of the polar retraction.
This work provides C++20 AVX-512 and Triton GPU kernel implementations, and was tested in 600 stress trials.
The geometric axioms of this work were formally verified in Lean 4.
The core metrics (orthogonality bound, sub-millisecond latency, 600 stress trials) are all reported by the authors and have not yet been independently reproduced.
All Facts (7)
This orthogonality error bound (10^-15) is the part the authors actively invite community criticism on, and the authors explicitly state that they hope to receive challenges regarding the boundaries of the polar retraction.
50%UnverifiedThe core claim of POLYDIM is: rather than using discrete one-dimensional autoregressive tokens to carry reasoning, it maps reasoning trajectories directly onto isometric geodesics on the sphere S^(D-1), where D ≥ 1024.
50%UnverifiedThe paper optimizes on the Stiefel manifold St(k, D), using a fourth-order polar decomposition retraction together with a Neumaier compensated accumulator, and reports an orthogonality error of ||X^T X - I_k|| ≤ 10^-15 with no catastrophic drift.
50%UnverifiedThe core metrics (orthogonality bound, sub-millisecond latency, 600 stress trials) are all reported by the authors and have not yet been independently reproduced.
50%UnverifiedThis work provides C++20 AVX-512 and Triton GPU kernel implementations, and was tested in 600 stress trials.
50%UnverifiedThe geometric axioms of this work were formally verified in Lean 4.
50%UnverifiedA research group on Reddit has publicly released their work POLYDIM / EinsofOS, which attempts to circumvent the quadratic complexity of attention and the information loss of one-dimensional tokens at the representation level.
50%