Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning Task-Sufficient World Models by Synergizing Agentic Exploration and Structured Modeling
Fan Feng, Yujia Zheng, Minghao Fu, Yongqiang Chen, Guangyi Chen, Kevin Patrick Murphy, Biwei Huang, Kun Zhang
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 MIST-WM latent masking with agentic probing instead of random data collection and unstructured latents (ablation) | MIST-WM · RoboSuite Stacking probing | average R2: 0.31 -> 0.78 | Improved |
| collect data with METRA skills instead of random (ablation) | world model data curation · RoboSuite Stacking probing | average R2: 0.31 -> 0.59 | Improved |
| use structured representation learning (I-Factor, Denoised MDP) instead of DINO-WM (ablation) | world model backbone · RoboSuite Stacking probing | average R2: 0.40 -> 0.48 / 0.56 | Improved |
| add MI term in the minimality score (ablation) | MIST-WM · single-task and generalization (normalized to Oracle) | normalized score | Improved |
| add masking in the minimality score (ablation) | MIST-WM · generalization tasks | normalized score | Improved |
| add sufficiency term (ablation) | MIST-WM · single-task and generalization | normalized score | Improved |
| add adaptive curriculum instead of random task order (ablation) | MIST-WM · generalization vs single-task | normalized score | Improved |
| replace Dreamer-v3 encoder with DINO-WM or R3M features (ablation) | MIST-WM · ablation tasks | normalized score | Mixed |
| use MIST-WM instead of TD-MPC2 | DMControl single-task · Reacher-Hard | episodic reward | Hurt |
| add active probing skills with random-vs-curiosity data (ablation) | MIST-WM exploration · generalization tasks | normalized score | Mixed |
| add MIST-WM structured modeling on top of Dreamer-V3 | Dreamer-V3 · training cost | GPU days: 0.25 -> 0.35 (DMControl) | Hurt |
Closest papers by meaning
- Modeling Unseen Environments with Language-guided Composable Causal Components in Reinforcement Learning (ICLR 2025 · similarity 0.89)
- RIG: Synergizing Reasoning and Imagination in End-to-End Generalist Policy (ICLR 2026 · similarity 0.89)
- General agents need world models (ICML 2025 · similarity 0.89)
- Dream to Manipulate: Compositional World Models Empowering Robot Imitation Learning with Imagination (ICLR 2025 · similarity 0.89)
- Dynamics-Aligned Latent Imagination in Contextual World Models for Zero-Shot Generalization (NeurIPS 2025 · similarity 0.88)
- RoboDreamer: Learning Compositional World Models for Robot Imagination (ICML 2024 · similarity 0.88)
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.
