Predictive but Not Plannable: RC-aux for Latent World Models
TLDR
RC-aux corrects spatiotemporal mismatch in latent world models for better long-horizon planning with minimal cost.
Reasoning
The paper clearly identifies a key limitation of latent world models and proposes a lightweight auxiliary objective to improve planning alignment. Strengths include a focused problem statement and empirical validation on multiple tasks, but the approach is limited to reconstruction-free models and may not generalize broadly.
Read-first score
Read-first score 54.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 25.
Field roles
Rank sensitivity
Stability: volatile; rank range: 465.
Keyword Scores
Deep Analysis
Innovations
- RC-aux (Reachability-Correction auxiliary objective) to address spatiotemporal mismatch in latent world models
- Multi-horizon open-loop prediction along the time axis for planning-aligned supervision
- Budget-conditioned reachability supervision with temporal hard negatives along the space axis
- Reachability-aware planner at test time that favors goal-directed and attainable trajectories
Methodology
RC-aux is a lightweight auxiliary objective added to reconstruction-free latent world models, keeping the backbone unchanged. It adds planning-aligned supervision via multi-horizon open-loop prediction and budget-conditioned reachability supervision with temporal hard negatives. The method is instantiated on LeWorldModel and evaluated under continuation-training and matched-from-scratch settings on goal-conditioned pixel-control tasks and a LIBERO-Goal extension.
Key Results
RC-aux improves LeWM-style planning with modest additional cost, suggesting that planning with latent world models depends on the representation encoding temporal and geometric structure beyond predictive accuracy.