Research map / Reinforcement learning / Policy gradient and actor-critic
Horizon Imagination: Efficient On-Policy Rollout in Diffusion World Models
Lior Cohen, Ofir Nabati, Kaixin Wang, Navdeep Kumar, Shie Mannor
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add stable action sampling instead of naive per-step sampling (ablation) | Horizon Imagination · Atari100K and Craftium control (Boxing, Gopher) | episodic return | Improved |
| reduce denoising budget from 32 to 16 with decay horizon 4 (ablation) | Horizon Imagination · Atari100K and Craftium control | episodic return: B=32 -> B=16 | No clear effect |
| use decay horizon 4 instead of autoregressive (k=1) at B=32 (ablation) | Horizon Imagination · Craftium/ChopTree-v0 | episodic return | Improved |
| halve denoising budget from 32 to 16 | Horizon Imagination · Atari100K training time | training time per run (hours): 27 hours -> 19 hours | Improved |
| halve denoising budget from 32 to 16 | Horizon Imagination · Runtime (RTX 4090) | controller training step (ms): 1871 -> 1071 | Improved |
| use parallel decay horizon 4-16 instead of autoregressive at low to medium budgets (ablation) | Horizon schedule · World model generation quality | FVD | Improved |
| use parallel decay horizon at extreme budgets B>=128 (ablation) | Horizon schedule · World model generation quality | FVD | Hurt |
| increase denoising budget (ablation) | Horizon schedule · World model generation quality | MSE vs ground truth | Hurt |
| use Pyramidal schedule instead of Horizon schedule at high budgets (ablation) | Horizon schedule · World model generation quality | FVD | Hurt |
Closest papers by meaning
- Open-World Reinforcement Learning over Long Short-Term Imagination (ICLR 2025 · similarity 0.86)
- Diffusion for World Modeling: Visual Details Matter in Atari (NeurIPS 2024 · similarity 0.86)
- Dynamics-Aligned Latent Imagination in Contextual World Models for Zero-Shot Generalization (NeurIPS 2025 · similarity 0.85)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.85)
- Stochastic Control for Fine-tuning Diffusion Models: Optimality, Regularity, and Convergence (ICML 2025 · similarity 0.85)
- Score Regularized Policy Optimization through Diffusion Behavior (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
