Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning to Predict Without Looking Ahead: World Models Without Forward Prediction
Daniel Freeman, David Ha, Luke Metz
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 |
|---|---|---|---|
| Use observational dropout with peek probability p=10% | Joint policy and world-model training on swing-up cart-pole · Swing-up cart-pole | cumulative score: higher than ∼ 500 | Improved |
| Use observational dropout with peek probability p=5% | Joint policy and world-model training on swing-up cart-pole · Swing-up cart-pole | task success rate: solves the task half of the time | Mixed |
| Train the world model jointly with the policy using observational dropout rather than a forward-predictive loss | Swing-up cart-pole transfer, 120-unit models · Swing-up cart-pole transfer | cumulative score | Improved |
| Increase the forward-predictive model width by 10x | Forward-predictive cart-pole baseline · Swing-up cart-pole transfer | cumulative score: 10x | No clear effect |
| Use a convolutional instead of a fully connected world model | Observational-dropout grid-world agent · Grid world | reward: 10650 parameters; 1201 learnable parameters | Improved |
| Use a world model trained at lower peek probabilities | Linear policy using frozen world-model features · Car Racing | score: around 40% | Improved |
| Reduce peek probability too close to zero | Joint policy and world-model training on Car Racing · Car Racing | score | Hurt |
| Remove VAE noise by using the encoder's predicted mean as the latent observation (ablation) | Car Racing agent · Car Racing | score | Improved |
| Replace the convolutional world model with a fully connected world model (ablation) | Observational-dropout grid-world agent · Grid world | reward: 10650 parameters; 1201 learnable parameters | Hurt |
| Train with a peek probability close to p=50% | Convolutional grid-world world model · Grid world | next-frame prediction correlation: p=50% | Improved |
Closest papers by meaning
- General agents need world models (ICML 2025 · similarity 0.88)
- Forecasting in Offline Reinforcement Learning for Non-stationary Environments (NeurIPS 2025 · similarity 0.87)
- Beyond Prediction: Steering VLM Agents with Retrospective World Modeling (NeurIPS 2026 · similarity 0.87)
- Simplifying Latent Dynamics with Softly State-Invariant World Models (NeurIPS 2024 · similarity 0.86)
- Inverting the Bellman Equation: From $Q$-Values to World Models (NeurIPS 2026 · similarity 0.86)
- Continual Reinforcement Learning by Planning with Online World Models (ICML 2025 · similarity 0.86)
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.
