World Models with Configurable Perception
To address this problem, we introduce ConfigWM, which realizes configurable perception through task-conditioned state extraction.
ProofPaper ↗
Key points
- Effective world modeling and planning need to abstract observations into world state representations while filtering out irrelevant information.
- Given a language description of the task at hand, ConfigWM uses it as a configuration signal to steer how visual embeddings are mapped into a task-specific state space, selectively preserving information relevant to the current task while suppressing task-irrelevant variations.
- Action-conditioned dynamics are then learned in this task-specific state space for planning.
- Consistent gains in physical-quantity probing (+0.365 Pearson r) and distracted-to-clean retrieval (+75.5%) further indicate that ConfigWM learns states that preserve task-relevant information while remaining robust to distractors, supporting reliable world modeling and planning in noisy worlds.
Sources (2)
- [1]World Models with Configurable PerceptionHugging Face Daily Papers · Oct 20, 12:00 AM
To address this problem, we introduce ConfigWM, which realizes configurable perception through task-conditioned state extraction.
Effective world modeling and planning need to abstract observations into world state representations while filtering out irrelevant information.
- [2]World Models with Configurable PerceptionarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 09:26 AM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026EmbodiedWorld-200K: Toward Embodied World Agents via Embodied-Planning Dataset and Interactive World Models
- Oct 19, 2026Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
- Oct 18, 2026The Explainability Illusion: Must ViTs Exhibit Patch-Decoding Interpretability?
- Oct 18, 2026Comfy-Org/Fibo-2.0
- Oct 18, 2026Google Open-Sources Harness2, a Recursive Agent Harnessing Framework
- Oct 18, 2026Elevatormusic/DeepSeek-V4.1-Flash-1017-EXL3-TensorFold-2x-DGX-Spark: DeepSeek-V4.1-Flash-1017 EXL3 2.9bpw on 2x NVIDIA DGX Spark (GB10) with TensorFold: 2.1-2.3x the vLLM kit's decode speed.