Research map / Reinforcement learning / Visual and representation RL
Compositional Planning with Jumpy World Models
Jesse Farebrother, Matteo Pirotta, Andrea Tirinzoni, Marc G Bellemare, Alessandro Lazaric, Ahmed Touati
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 compositional planning over GHMs to a base policy | GC-BC · OGBench cube-2 | success rate: 0.15 -> 0.97 | Improved |
| Add compositional planning over GHMs to a base policy | CRL · OGBench cube-4 | success rate: 0.00 -> 0.39 | Improved |
| Add compositional planning over GHMs to a base policy | HFBC · OGBench antmaze-giant | success rate: 0.42 -> 0.79 | Improved |
| Add compositional planning over GHMs to a base policy | GC-TD3 · OGBench antmaze-medium | success rate: 0.65 -> 0.65 | No clear effect |
| Add compositional planning over GHMs to a base policy | HFBC · OGBench antmaze-medium | success rate: 0.94 -> 0.94 | No clear effect |
| Plan over policy sequences (CompPlan) instead of generalized policy improvement (ablation) | CompPlan · OGBench long-horizon domains | relative success rate improvement: 89% relative improvement over GPI | Improved |
| Plan over policy sequences with jumpy models instead of action-level planning with a one-step world model (ablation) | CompPlan · OGBench long-horizon domains | relative success rate improvement: 201% gain over ActionPlan | Improved |
| Use test-time CompPlan with HFBC instead of learned hierarchies | CompPlan · OGBench cube-4 | success rate: 0.09 -> 0.67 vs SHARSA, 0.00 -> 0.67 vs HIQL | Improved |
| Use CompPlan with HFBC instead of HIQL | CompPlan · OGBench antmaze-medium | success rate: 0.96 -> 0.94 | Hurt |
| Add horizon consistency loss (TD-HC) to GHM training (ablation) | GHM with GC-1S · OGBench antmaze-giant | EMD: 7.29 -> 5.25 (28% reduction) | Improved |
| Add horizon consistency loss (TD-HC) to GHM training (ablation) | GHM with CRL · OGBench antmaze-giant | EMD: 6.77 -> 5.74 | Improved |
| Add horizon consistency loss (TD-HC) to GHM training (ablation) | GHM with CRL · OGBench cube-1 | EMD: 1.60 -> 1.57 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- HIQL — HIQL: Offline Goal-Conditioned RL with Latent States as Actions (NeurIPS 2023) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Modeling Unseen Environments with Language-guided Composable Causal Components in Reinforcement Learning (ICLR 2025 · similarity 0.88)
- Beyond Prediction: Steering VLM Agents with Retrospective World Modeling (NeurIPS 2026 · similarity 0.87)
- Predictive but Not Plannable: RC-aux for Latent World Models (NeurIPS 2026 · similarity 0.87)
- Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning (NeurIPS 2026 · similarity 0.87)
- GenPlan: Generative Sequence Models as Adaptive Planners (AAAI 2025 · similarity 0.87)
- Temporal Representation Alignment: Successor Features Enable Emergent Compositionality in Robot Instruction Following (NeurIPS 2025 · similarity 0.87)
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.
