The paper introduces HamiFormer, a learned physical simulator that combines whole-window denoising with residual-corrected Hamiltonian propagation. A key component is Parallel Local Affine Scan (PLAS), which uses affine symplectic maps to solve implicit Hamiltonian relations efficiently. The model also employs a Regime Model Tree to balance accuracy for typical states against large tail errors.
In 192-step evaluations on two synthetic benchmarks (HamiBalls-1 and HamiBalls-2), HamiFormer achieves normalized phase-space MSE reductions of 26.3% over PhysiFormer and 21.4% over DiT, respectively, with comparable model capacities. PLAS demonstrates lower solver error and runtime compared to sequential explicit Euler. The method also provides theoretical conditions for physically consistent refinement and finite-window error bounds.
This work is interesting because it addresses the long-standing problem of error accumulation in autoregressive physical simulators by integrating local Hamiltonian dynamics with whole-window diffusion refinement. The dual-expert approach with mixed-state feedback offers a principled way to attenuate inherited prediction errors, potentially improving reliability in long-horizon trajectory prediction.