HaM-World: Soft-Hamiltonian World Models with Selective Memory for Planning

arXiv 2026 56.1 method

TLDR

HaM-World combines Hamiltonian dynamics with selective memory (Mamba) for stable long-horizon planning in model-based RL, achieving strong results on DMC tasks and OOD perturbations.

Reasoning

Strengths include novel integration of Hamiltonian mechanics and state-space memory to address instability in world model rollouts, with thorough empirical validation on multiple tasks and OOD conditions. Weaknesses are limited to simulated control tasks and lack of analysis on more complex or high-dimensional environments.

Read-first score

Read-first score 56.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 29%

72.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Citation impact 18%

72.8

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.72807024

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,metric

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 403.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

generative world model

interactive world model

world simulator

video world model

Deep Analysis

Innovations

Decomposition of latent state into canonical (q, p) subspace and context subspace c
Use of Mamba selective state-space memory as history-conditioned input to latent dynamics
Hamiltonian vector field with learnable residual/control dynamics for (q, p) evolution
Soft-Hamiltonian dynamics design that bounds action-free Hamiltonian-energy drift
Shared latent state for dynamics prediction, reward/value estimation, imagined rollouts, and CEM action search

Methodology

HaM-World decomposes the latent state into a canonical (q, p) subspace and a context subspace c, using Mamba selective state-space memory as history-conditioned input. The (q, p) evolves via an energy-derived Hamiltonian vector field plus learnable residual/control dynamics, while c captures semantic, dissipative, and non-conservative factors. The model is evaluated on four DeepMind Control Suite tasks, with baselines including a strong baseline model, using metrics such as Avg. AUC, long-horizon rollout error, MSE cells, and OOD return.

Key Results

HaM-World achieves the highest Avg. AUC (117.9, +9.5%), reduces long-horizon rollout error to 45% of a strong baseline, and wins 11/12 k in {3,5,7} MSE cells. Under 12 OOD perturbations, it achieves the highest return in every condition, with average OOD-return gains of 10.2% on Finger Spin and 13.6% on Reacher Easy.

Resources

Paper