Research map / Video / Action recognition
Factored Latent Action World Models
Zizhao Wang, Chang Shi, Jiaheng Hu, Kevin Rohling, Roberto Martín-Martín, Amy Zhang, Peter Stone
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 causal temporal attention to the factorizer (ablation) | FLAM · MultiGrid | PSNR: 51.6 -> 56.5 | Improved |
| Add causal temporal attention to the factorizer (ablation) | FLAM · MultiGrid (DCI) | disentanglement: 0.33 -> 0.91 | Improved |
| Add causal temporal attention to the factorizer (ablation) | FLAM · MultiGrid | FVD: 36.6 -> 2.9 | Improved |
| Condition per-slot IDM on own next slot only instead of all slots (vs Global-Coupled) (ablation) | FLAM · MultiGrid | PSNR: 56.3 -> 56.5 | No clear effect |
| Use per-slot latent actions instead of Global-Coupled variant (ablation) | FLAM · MultiGrid (DCI) | disentanglement: 0.79 -> 0.91 | Improved |
| Increase number of factors K from 2 to 4 (matching 4 agents) (ablation) | FLAM · MultiGrid 4 agents | PSNR: 53.4 -> 56.5 | Improved |
| Increase number of factors K from 4 to 16 (ablation) | FLAM · MultiGrid 4 agents | PSNR: 56.5 -> 56.5 | No clear effect |
| Increase KL coefficient beta from 1e-3 to 1e-2 (ablation) | FLAM · MultiGrid | PSNR: 56.3 -> 24.9 | Hurt |
| Use factored latent actions (FLAM) instead of single monolithic latent action (AdaWorld) | AdaWorld · Average over 5 datasets | PSNR: 24.7 -> 34.9 | Improved |
| Use FLAM instead of Genie | Genie · Average over 5 datasets | FVD: 717.5 -> 419.6 | Improved |
| Use FLAM instead of PlaySlot (joint instead of two-stage object-centric training) | PlaySlot · Average over 5 datasets | PSNR: 29.1 -> 34.9 | Improved |
| Use FLAM instead of ground-truth-action World Model | World Model · Bigfish | PSNR: 23.9 -> 30.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- PlaySlot — PlaySlot: Learning Inverse Latent Dynamics for Controllable Object-Centric Video Prediction and Planning (ICML 2025) · better on 4 of 4 shared results
- SlotFormer — SlotFormer: Unsupervised Visual Dynamics Simulation with Object-Centric Models (ICLR 2023) · better on 4 of 4 shared results
- AdaWorld — AdaWorld: Learning Adaptable World Models with Latent Actions (ICML 2025) · 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
- Latent Action Learning Requires Supervision in the Presence of Distractors (ICML 2025 · similarity 0.88)
- PlaySlot: Learning Inverse Latent Dynamics for Controllable Object-Centric Video Prediction and Planning (ICML 2025 · similarity 0.88)
- What Do Latent Action Models Actually Learn? (NeurIPS 2025 · similarity 0.88)
- LAFP: Preserving Latent Action Structure in Latent Policy Learning via Flow Matching (NeurIPS 2026 · similarity 0.87)
- SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation (ICLR 2025 · similarity 0.87)
- Learning Visual Feature-Based World Models via Residual Latent Action (NeurIPS 2026 · 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.
